Journal Pre-proof

Data-driven Models for Predicting Compressive

Strength of 3D-printed Fiber-Reinforced Concrete
using Interpretable Machine Learning Algorithms

Muhammad Arif, Faizullah Jan, Aïssa Rezzoug,

Muhammad Ali Afridi, Muhammad Luqman,
Waseem Akhtar Khan, Marcin Kujawa, Hisham
Alabduljabbar, Majid Khan

PII: S2214-5095(24)01086-6
DOI: https://doi.org/10.1016/j.cscm.2024.e03935
Reference: CSCM3935

To appear in: Case Studies in Construction Materials

Received date: 4 May 2024
Revised date: 9 October 2024
Accepted date: 30 October 2024
Please cite this article as: Muhammad Arif, Faizullah Jan, Aïssa Rezzoug,
Muhammad Ali Afridi, Muhammad Luqman, Waseem Akhtar Khan, Marcin
Kujawa, Hisham Alabduljabbar and Majid Khan, Data-driven Models for
Predicting Compressive Strength of 3D-printed Fiber-Reinforced Concrete using
Interpretable Machine Learning Algorithms, Case Studies in Construction
Materials, (2024) doi:https://doi.org/10.1016/j.cscm.2024.e03935
This is a PDF file of an article that has undergone enhancements after acceptance,
such as the addition of a cover page and metadata, and formatting for readability,
but it is not yet the definitive version of record. This version will undergo
additional copyediting, typesetting and review before it is published in its final
form, but we are providing this version to give early visibility of the article.
Please note that, during the production process, errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
© 2024 The Author(s). Published by Elsevier Ltd.
 Data-driven Models for Predicting Compressive Strength of 3D-printed Fiber-
Reinforced Concrete using Interpretable Machine Learning Algorithms

Muhammad Arif1, Faizullah Jan2,*, Aïssa Rezzoug3,*, Muhammad Ali Afridi4, Muhammad
Luqman1, Waseem Akhtar Khan4, Marcin Kujawa2, Hisham Alabduljabbar5, Majid Khan6,*

1
Department of Civil Engineering, University of Engineering and Technology, Peshawar,
25120, Pakistan
2
Gdańsk University of Technology, Faculty of Civil and Environmental Engineering, Poland

of
3
Imam Mohammad Ibn Saud Islamic University (IMSIU), Riyadh, Saudi Arabia
4
Department of Civil Engineering, University of Louisiana at Lafayette, Lafayette, LA, 70503,

ro
USA
5
Department of Civil Engineering, College of Engineering in Al-Kharj, Prince Sattam Bin
Abdulaziz University, Al-Kharj 11942, Saudi Arabia
6
-p
Department of Civil Engineering, Southern Illinois University Edwardsville, Edwardsville,
re
IL, 62026, USA
*
Corresponding author: [email protected] (F.J), [email protected] (A.R),
lP

[email protected] (M.K)
a
rn

Abstract: 3D printing technology is growing swiftly in the construction sector due to its
numerous benefits, such as intricate designs, quicker construction, waste reduction,
u

environmental friendliness, cost savings, and enhanced safety. Nevertheless, optimizing the
Jo

concrete mix for 3D printing is a challenging task due to the numerous factors involved,
requiring extensive experimentation. Therefore, this study used three machine learning
techniques, including Gene Expression Programming (GEP), Multi-Expression Programming
(MEP), and Decision Tree (DT), to forecast the compressive strength of 3D printed fiber-
reinforced concrete (3DP-FRC). The dataset comprises 299 data points with sixteen variables
gathered from experimental research studies. For training the model, 70% of the dataset was
used, while the remaining 30% was reserved for model testing. Several statistical metrics were
utilized to evaluate the accuracy and applicability of the models. In addition, SHapley Additive
exPlanations (SHAP), partial dependence plots, and individual conditional expectations
approach were employed for the interpretability of the models. The proposed GEP, MEP, and
DT models indicated enhanced efficacy, exhibiting correlation coefficient (R) scores of 0.996,
0.987, and 0.990, with mean absolute errors (MAE) of 1.029, 4.832, and 2.513, respectively.
Overall, the established GEP model demonstrated exceptional performance compared to MEP
and DT, showcasing high prediction precision in assessing the strength of 3DP-FRC. Moreover,
a simple empirical formulation has been devised using GEP to predict the compressive strength,
offering a simplified and efficient approach for predicting 3DP-FRC strength. The SHAP
approach identified water, silica fume, fiber diameter, curing age, and loading directions as
leading controlling parameters in predicting strength of 3DP-FRC. In summary, the proposed
models can potentially minimize both the computational workload and the need for
experimental trials in formulating the mixed design of 3D-printed concrete.

of
Keywords: Machine learning; 3D-printed fiber reinforced concrete; model interpretability,

ro
compressive strength

1. Introduction -p
Since the turn of the millennium, Additive Manufacturing (AM) has evolved alongside
re
digitization, technological advancements, and modernization [1]. This groundbreaking
development has encouraged automated techniques, notably in the Construction Industry (CI),
lP

where extrusion-based 3D has become prevalent. Current challenges, including the lack of
trained labor, resource exhaustion, and protection concerns, have hindered the progress of the
a

CI amid the tremendous advancements in urbanization and industrialization [2], [3]. 3D

rn

printing (3DP) technology gives construction researchers and engineers a unique perspective
by stacking printed material planes to form a 3D design model. This technique seeks to
u

substantially diminish solid waste creation in the CI, bringing cost savings and reducing skilled
labor needs [4]. Recent research demonstrates that 3DP technology in the construction sector
Jo

has the potential to reduce 30–60% of waste, 50–80% of the workforce, and 50–70% of
production time; thus, applications of 3D-printed concrete in civil engineering are modern
fundamentals and consequently, construction automation is inevitable to be one of the dominant
trends in the foreseeable future of CI [5].
AM technology enables the creation of distinctive architectural elements or large-scale building
components using a layer-by-layer method, eliminating the need for formwork. Furthermore,
various 3D printing concrete methods have been developed to integrate AM standards into
concrete production [5], [6]. Two classes divide these technologies into powder-based
(Emerging Objects, D-shape technique) and extrusion-based (Contour Crafting (CC), Concrete
Printing, CONPrint3D) groups [7], [8], [9], [10]. Other novel techniques recently presented are
rock printing, Fused Deposition Modeling (FDM), and MultiJet fusion [11], [12]. However,
only for concrete and earthen architecture, this approach's potential to create large-scale
structures without formworks [13] has proven essential to the design process regardless of the
productivity drop since it saves money and material (Fig. 1).

of
ro
-p
re
lP

Fig. 1. Different Kinds of Construction [14]

As 3D printing technology develops, desired material properties must be attainable to satisfy

the technology's high-efficiency requirements. The concrete researchers have examined the
a

mechanical characteristics [9], [15], [16], [17], [18], [19] and workability [1], [20], [21], [22],
rn

[23], [24], [25], [26] of 3D-printed concrete, as well as the relationships between them [27].
Additionally, it was not possible to determine the ideal mix percentage and a technique for
u

assessing the printability of 3DPC [28]. Researchers investigated a wide range of constituents,
Jo

including mineral powders [25] and geopolymer-based materials [16], [18], [29], to produce
the most effective 3DPC and conventional cement-based materials.

Although 3DPC infrastructure meets international dynamic standard procedures, several

structures have been 3D printed, but only some are made practical, owing to emerging safety
concerns. For instance, in 2015, Eindhoven University of Technology in the Netherlands
printed the first multistory building [30], which complied with all applicable standard criteria.
Furthermore, in 2016, the local administration executed the initial 3D-printed office in Dubai,
which was equally successful as a nominal building [31]. The Institute of Advanced
Architecture of Catalonia (IAAC) built the first 3D-printed pedestrian bridge spanning 40ft
using the D-shape printing technique for all eight bridge components. The subsequent year,
[32] printed and tested a prototype bridge in six pieces using the CC technique to accommodate
the printer's span limitations successfully. On-site well-developed testing procedures
ascertained its structural strength and integrity without pre-established standard criteria for
evaluation of printed structure. 3DPC technology has the potential capacity to create, fabricate,
repair, or restore complex structures in circumstances where conventional formwork offers
difficulties [31], [33]. 3DPC structures have far-reaching potential despite the limitations of 3D
technology, such as the complication of the process, absence of specifications, and fiscal
viability [34]. Due to the versatility of this innovation across numerous industries, construction
processes on Mars and the Moon may become streamlined. As a consequence, the goal to
colonize extraterrestrial regions could potentially accelerate [35]. As the technology of 3DP
advances, its potential and advantages become more and more evident.

of
The construction industry acknowledges the potential of 3DPC, but conventional concrete is

ro
prone to brittle failure and requires reinforcement to fulfill essential construction criteria such
as elasticity, creep, and durability [30], [36], [37], [38]. Furthermore, 3DPC exhibits less
-p
strength than regular cast concrete due to factors such as the lack of reinforcing, mix
compaction, and concerns with uniformity as it achieved layer upon layer [39]. As a result,
re
structural components produced with 3DPC may perform less than those made using traditional
techniques. Hence, it is essential to strengthen the 3DPC, and enhancement in material
lP

attributes is required to fulfill the performance criteria [40]. 3DP technology can provide an
innovative way to reinforce concrete by printing steel reinforcement or steel tubes[41].
a

Nevertheless, associated with fiber-reinforced 3D printed concrete, this technique is still in the
rn

early stages of exploration. A practical approach to the reinforcing difficulty involves

substituting the steel bars with shortened fibers[42]. Fibers, mainly consisting of carbon [43],
u

glass[4], basalt [4], [44], polyethylene (PE) [45], polypropylene (PP) [46], polyvinyl alcohol
(PVA) [47], and steel fibers [48] are among them. However, there is a pressing requirement for
Jo

fiber-reinforced cement-based materials to expand the use of 3DP concrete in creating high-
performance structures[49], [50].

However, incorporating reinforcement into the process of 3D concrete printing offers

challenges due to its incremental layering technique. Although researchers have not yet
established a completely automated system for reinforcing in 3D concrete printing, prior
studies have proposed several strategies to integrate reinforcement into 3DPC [51], [52]. A
primary method for strengthening 3D-printed building components involves depositing
materials layer by layer around manually inserted rebars [53]. An alternative method proposes
incorporating steel cable into the extruded layers [54]. Additionally, the researchers
investigated other forms of reinforcement, including fabrics, recycled glass, and wire meshes,
that may be included in the extruded layers [6], [45], [55]. Although implementing
reinforcement during layer deposition necessitates electromechanical units alongside
additional subsystems, designing and improving automated processes for these activities is
essential to ensure an uninterrupted building process. Furthermore, current reinforcing
techniques may add complexity to manufactured 3DPC, compromising its advantage of
automated building [56].

Both the curing age and curing method affect the strength of 3DP-FRC. The strength of
composite 3DP-FRC is enhanced using curing-based active and passive techniques, which
lowers the nano deformation cracks [57]. Passive techniques include using a well-suited

of
mixture of materials, including anti-shrink fibers like glass, polypropylene, polyethylene,
basalt, or steel[57]. This method may also use interior curing techniques, such as the addition

ro
of shrinkage-reducing agents (SRA) or utilizing fine aggregates soaked in water[58], [59].
However, active approaches reduce water evaporation from the material's surface. Some of
-p
these methods are moistening the concrete, using anti-evaporation chemicals (membrane
waterproofing) or foil, and protecting it from direct sunlight and wind [60]. The diameter of
re
fibers also positively affects the flexure capacity of 3DFRC in contrast to fiber length, which
sometimes renders the CS. It gives researchers the idea of using an appropriate aspect ratio of
lP

fibers relative to nozzle diameter instead of using any single geometrical fiber property[61].
[62] investigated the effect of water content on 3D printed graphene concrete with water-binder
a

ratio variation from 0.25-0.5. The pore spheroid of 3DPFC increases with the same concrete
rn

strength as the amount of silica fume in the material increases [63].

To quantify the mechanical performance of 3DP-FRC composites, researchers are examining

u

compressive strength (CS) as a significant property. The highest CS of any 3DP-FRC, reaching
Jo

159 MPa, is shown by 3DP steel fiber reinforced concrete[48]. The anisotropic characteristics
of 3DPFRC with fiber amount changes up to 1% identified material anisotropy for loading
orientations of 90o, 45o, and 0o relative to the printing direction [39]. As more material is cast
longitudinally during printing, the compressive strength in the printing direction is always the
highest. Singh et al. [39] examined the anisotropic properties of 3DPFRC with fiber quantity
changes up to 1% and found CS decreased by 3-14% at 0o compared to 45o and 12-26% at 90o.
The compressive strength of molded specimens reinforced with PE fibers was lower than that
of printed mixtures[64]. Adding 1% glass fibers to the printed mixtures resulted in a notable
108% increase in strength [65]. With more fibers, achieving the desired porosity of bulk
slurry[66] composite from fiber orientation and winding can reduce strength[67]. PVA fibers
reduced compressive strength by 1.5%, like glass and basalt fibers [55]. The horizontally
reinforced specimen often breaks along the soft surface during axial compression, exposing the
horizontal reinforcement. With increases of up to 2% in volume fraction, including 3DPC with
steel fiber significantly enhanced CS and FS compared to conventional cast specimens [48].
These results demonstrate that enhancing the 3D concrete mixture with a suitable dosage of
fibers greatly boosts CS, particularly with steel fibers; it shows remarkable efficiency [48],
[68].

The primary variables affecting the CS of 3DP-FRC are intrinsic and extrinsic. Intrinsic factors
include the mix percentage, raw materials, kind, quantity, and geometrical characteristics of
the fiber, while experimental extrinsic factors include the curing age and loading direction. It

of
has been noted that there is currently no universally accepted standard, criteria, or procedure
in this field. This lack is seen as a challenge, as it adds complexity to the process of identifying

ro
the optimal mix design for enhancing the properties of 3DP-FRC. In addition, investigating the
complex relationship between the input features and the mechanical performance of 3DP-FRC
-p
using experimental tests is expensive and time-consuming. One possible solution is using data-
driven techniques, for example, machine learning (ML) algorithms, to forecast the mechanical
re
performance of 3DP-FRC with specified parameters.

To estimate the mechanical characteristics of various fiber-reinforced concrete (FRC) and other
lP

cementitious products, such as PVA-FRC [69], [70], PP-FRC [71], PE-FRC [72], basalt FRC
[44], [73], steel FRC [74], FRP-CBS [75], FRP-RC [76], HFR-SCC [77], SRFC [78], SFRC-
a

SWS [79], SRER [80] researchers built a variety of machine learning (ML) models, listed in
rn

Table 1. The ML models were also extensively used to address diverse technological
challenges; some are very reliable. For instance, Light Gradient Boosting Machine
u

(LightGBM) outperformed ICA-XGBoost [81] in predicting the CS of recycled aggregate

Jo

concrete. According to [82], CatBoost demonstrated greater accuracy than XGBoost [83],
Support Vector Machine (SVM), and LighGBM in forecasting the strength of concrete-filled
axial steel tubular columns.

Table 1. Previous studies on predicting CS of different types of concrete

Concrete type Method Dataset R2 Ref.
HPC-GGBS LightGBM 152 0.96 [42]
HSC GEP 357 0.91 [43]
HPC-GGBS LightGBM 152 0.96 [42]
PCM XGBoost 154 0.96 [4]
ECC ANN 79 0.99 [70]
 HPC RF 190 0.98 [45]
MSC NGBoost 754 0.98 [41]
Normal concrete SVM 15 0.95 [47]
RHAC ANN 192 0.9 [46]
UHPC LightGBM 113 0.99 [70]
RAC ICA-XGBoost 209 0.98 [84]
FAC SVM 270 0.98 [85]
Geopolymer concrete RF 35 0.93 [48]
RAC LightGBM 515 0.96 [74]
Normal concrete AdaBoost 1030 0.99 [38]

of
UHPC XGBoost 372 0.9 [73]

ro
In cementitious materials research, ML methods have been widely used to predict properties.
-p
However, there is a noticeable absence in the current literature regarding the wide utilization
of various ML models to evaluate the CS of 3DP-FRC. Limited research studies in the current
re
literature concentrate on forecasting the CS of 3DP-FRC using various ML models. Certain
studies have been used to forecast the strength of 3DP-FRC utilizing ML methods which are
lP

opaque and cannot yield an empirical equation which is crucial in easy practical engineering
applications of ML. Yet a notable gap persists wherein empirical formulations for the CS of
a

3DP-FRC remain unexplored. To fill this gap, this work utilizes the transparent evolutionary
rn

algorithms GEP and MEP to provide a mathematical formulation for the CS of 3DP-FRC. The
primary objectives of the research are as follows: (1) to construct predictive models for the
u

compressive strength of 3DP-FRC using DT, GEP, and MEP, (2) to formulate an empirical
Jo

equation for the CS of 3DP-FRC using GEP, and (3) to employ SHAP, ICE, and PDP methods
for interpretability to gain insights into the fundamental reason behind the ML predictions.

2. Research methodology

2.1. Data collection

To develop dependable ML models for predicting the CS of 3DP-FRC, a total of 299 data
points were gathered from existing experimental studies [9], [15], [17], [18], [19], [22], [41],
[45], [46], [47], [48], [54], [81], [82], [84], [86], [87]. The modeling considered sixteen
parameters as input variables to understand the influence factors and the entire ML model of
CS prediction. These parameters included Ordinary Portland Cement (OPC), Sand (S), Water
to Binder ratio (W/B), Fly Ash (FA), Ground Slag (GS), Silica Fume (SF), Super Plasticizer
(SP), Hydroxypropyl Methylcellulose (HPMC), Water (W), Fiber Volume Fraction (VF),
Curing Age (CA), Load Direction (LD), Fiber Diameter (DF), Fiber Length (LF), Aspect Ratio
(AR) and Fiber Type (FT). The database for 3DP-FRC includes five distinct fiber types: PE
(designated as 1), steel (designated as 2), PVA (designated as 3), PP (designated as 4), and
basalt (designated as 5). The research shows that the strength of 2DPFRC is significantly
affected by applying load at different angles[19], [61]. Accordingly, the 3PD-FRC,
compressive strength testing often used Y, X, and Z loading orientations. Y and X denote
parallel and perpendicular to the printing direction of a single filament, respectively, whereas
Z stands for vertical orientation. These loading orientations are designated as 1 for X

of
orientation, 2 for Y direction, and 3 for Z direction.
Table 2 presents an overview of the statistical description of the input and output parameters

ro
of 3DP-FRC. Statistical indicators such as mean, standard deviation, minimum, maximum,
kurtosis, and skewness are included to facilitate the understanding of the data. The variables
-p
exhibit skewness and kurtosis values within the recommended ranges of ±3 and ±10,
respectively. Furthermore, it is essential to mention that SF has the highest positive correlation
re
(+0.94) with compressive strength among all the input parameters with CS, as shown in Fig. 2.
Next, the FD with a correlation coefficient of +0.49, SP, and GS with a correlation coefficient
lP

of +0.48, and curing age with a correlation coefficient of +0.43. In contrast, the correlation
between the (W/B) and CS is strongly inverse and significant (r = -0.84), closely followed by
a

a similarly negative correlation relation between W and CS (r = -0.82).

rn

Table 2. The statistical description of variables in the database

Parameter Unit Mean SD1 Min Max Kurtosis SK2 VIF OPN3
u

OPC Kg/m3 579.84 228.33 285.30 1112.30 -0.38 0.81 3.53 Input
Jo

W/B - 0.23 0.07 0.16 0.35 -0.76 0.70 5.83 Input

S Kg/m3 830.18 414.77 246.00 1902.00 -0.04 0.14 1.91 Input
GS Kg/m3 151.81 169.77 0 450.00 -1.77 0.29 1.31 Input
FA Kg/m3 288.50 411.64 0 1141.10 -0.75 0.96 2.72 Input
HPMC Kg/m3 0.74 1.22 0 3.80 0.93 1.47 1.63 Input
VF % 0.01 0.01 0 0.02 -1.28 0.29 3.52 Input
SP Kg/m3 6.79 5.05 0 20.00 -0.23 0.66 5.29 Input
LD X,Y,Z 2.10 0.85 0 3.00 -1.19 -0.33 1.00 Input
W Kg/m3 267.81 84.77 182.00 427.90 -1.51 0.24 6.49 Input
SF Kg/m3 165.41 132.17 0 377.80 -1.46 -0.07 4.62 Input
CA Days 24.65 8.36 1 28.00 2.69 -2.14 2.53 Input
LF mm 7.43 3.81 0 18.00 0.67 0.85 1.53 Input
DF µm 42.85 50.30 15.00 200.00 5.81 2.73 1.62 Input
FT - 2.38 1.13 1.00 6.00 0.35 0.80 2.53 Input
AR - 287.46 245.26 0 1200.00 5.85 2.22 1.44 Input
CS MPa 60.68 41.32 0 153.40 -1.11 0.14 - Output
1
Standard Deviation, 2Skewness, 3Operation
Furthermore, determining multicollinearity and interdependency in the considered dataset is an
essential stage in ML prediction modeling. Multicollinearity occurs when two or more
prediction model features correlate significantly. The presence of multicollinearity in the
dataset may arise from problems encountered during its development, such as poorly designed

of
experiments, restrictions in data processing skills, or excessive observational data. This results

ro
in less reliable projections, hence lowering their degree of precision. In such cases, the
algorithm struggles to distinguish each predictor's influence, resulting in inconsistent model

-p
output predictions [88]. One technique to discover multicollinearity in data is to examine
pairwise correlations using a correlation matrix. A Pearson's correlation (r) heat map (Fig. 2)
re
shows a correlation among the model variables. Generally, an r-value greater than 0.8 indicates
a risk of multicollinearity. It is evident that there is no risk of multicollinearity since the r-value
lP

between the majority of the input variables is less than 0.8. However, Pearson’s correlation
technique has limitations [89] since significant correlations between predictors may not always
a

indicate multicollinearity since these ideas are separate [90], [91]. Therefore, researchers
rn

widely use the VIF (Variance Inflation Factor)[92]. Relative to an ideal design with
uncorrelated (orthogonal) x-variables, the VIFk of the regression coefficient rk measures how
u

much the variance of rk rises due to collinearity. That is to say, it quantifies the extent to which
Jo

collinearity 'inflates' the variance of the regression coefficient. The value of the kth diagonal
element in the inverse correlation matrix of the variables is denoted by the VIF [93], which the
following Eq.(1) Error! Reference source not found. can express.
-1
VIFk =(1- Rk2 ) (1)

The coefficient of determination (Rk2) for the xk regression on each of the remaining regressors
is shown. The VIFk ranges from 1 (which represents noncorrelated coefficients from orthogonal
designs) to infinity (which, when Rk = 1, represents perfect correlation). It is suggested that the
connection was impacted by collinearity when a variable had a VIFk > 1. A VIFk >10 indicates
a strong correlation (collinearity) between the variables, while there is no set cutoff. Due to the
significant reduction in the standard error of the regression coefficient, often caused by strong
correlations in the data, the regression coefficient is calculated with considerable uncertainty.
The VIF and correlation values falling below the designated limit demonstrate that there is little
chance of multicollinearity in the dataset that was gathered as presented in Table 3 and Fig. 2.
Table 3. VIF analysis of the dataset
Variab OP HPM L A W/ C
S FA GS SF W VF DF SP LF FT
les C C D R B A
VIF 3.5 1.9 2.7 1.3 4.6 6.5 3.5 1.0 1.6 1.5 5.8 5.3 2.5 1.4 2.5
1.64
value 5 2 4 2 5 3 4 1 3 4 6 2 4 5 4

of
ro
-p
re
a lP
u rn
Jo

Fig. 2. Pearson's correlation (r) heat map

3. Model development, evaluation, and interpretability

3.1. Description of ML models

Three distinct machine learning techniques, DT, MEP, and GEP, are used in the current research
to establish predictive models for CS of 3DP-FRC.
3.1.1. Decision tree (DT)

One popular ML approach that can apply to both regression and classification problems is
decision trees [94]. The DT simplifies data analysis and insight extraction by breaking it down
into smaller data subsets. DT is used in ML because it can manually duplicate operational
decisions, similar to flowcharts. DT machine learning is a flexible method that utilizes input
data to produce decisions or predictions using a tree-like structure. Communicating model
output to humans is a major machine-learning challenge. Machine learning optimizes tasks
without human input, making model output descriptions difficult. A decision tree structure
simplifies model decision-making logic since each decision branch is visible. To build a

of
decision tree, the dataset should consist of subsets that are more homogeneous about the goal
variable. The classification and regression tree method (CART) builds decision trees

ro
representing all possible responses to a decision or issue under established parameters. The
term "decision tree" makes sense since it is structured similarly to a tree, with branches growing
-p
outward from a root node as the tree matures[95]. The dataset is divided into discrete segments
re
by the DT, and at each division, the decision tree evaluates the variation between the actual and
predicted values. The process is repeated when errors occur in the calculation of each partition
lP

feature, which results in the lowest fitness function being determined as the dividing point.

DT starts with a root node and expands on successive branches to construct a structure. A DT's
a

design has each inner node acting as a decision point determined by a particular characteristic,
rn

and each terminal node denotes a predicted value [96]. The decision tree may not always
provide a straightforward option. Instead, it may provide data scientists with options to make
u

good decisions. Fig. 3 shows the flowchart of the DT. A decision tree generates rules that result
Jo

in conclusions using its branches, where a leaf node represents a particular result [97] In this
architecture, two main types of distinguishable nodes are the decision node, which has branches
that facilitate different options, and the leaf node, which does not have branches and contains
the ultimate decision.
 of
Fig. 3. Schematic Illustration of Decision Tree

ro
3.1.2. Gene expression programming (GEP)

-p
GEP was developed by Candida Ferreira [98] and is an adaption of Genetic Programming
(GA), which is based on creating computer programs that employ gene combination and
re
modification to solve complicated problems. Using recombination or mutation, GEP chooses,
lP

multiplies, and integrates genetic variation from a population of models and solutions
according to suitableness. The primary distinction between the three algorithms, however,
varies according to the kinds of people, models, or solutions, as applicable. This soft computing
a

technique is vital in engineering because it thrives when other models fail or have limited
rn

capacities. Encoded in linear chromosomes of constant length, GEP is a vast genotyping system
made up of expression trees of various sizes and shapes. GP individuals are non-linear entities
u

of different sizes and forms, GA individuals are symbolic strings of fixed chromosomes, and
Jo

GEP individuals encrypt in GPs. GEP is an evolutionary algorithm based on genotype and
phenotype that builds large tree-like computer programs that resemble real species by changing
in size, shape, and composition. GEP's genes/tree system can explore all solution space routes
with increasing degrees of hierarchy, much like life's DNA system does. GEP involves the
Expression Tree (ET) displayed in Fig. 5 and chromosomes (Fig. 4), which are phenotype and
genotype, respectively.
 Fig. 4. Expression Tree of Chromosome

Because of its versatility, GEP use is prevalent in various problem fields, including pattern

of
recognition, symbolic regression, and mathematical optimization [90], [99]. Compared to other
GAs, GEP notably shows faster convergence [98], [100] with better interpretable results from

ro
its chromosomal-level genetic activities [101]. Fig.6 depicts the ET of the GEP chromosome
tree with LISP language for crossover and mutation.
-p
re
a lP
rn

Fig. 5. GEP chromosome tree with LISP language

u

Fig. 6 schematically shows the major phases of GEP modeling. The flowchart indicates that
Jo

GEP modeling starts by randomly generating chromosomes for predefined quantities, which
adhere to the Karva language (a language that demonstrates symbols) for introducing the
chromosomes. Onward, these chromosomes transform, and their effectiveness is evaluated
under predetermined fitness conditions [91], [102]. Then, some chromosomes carry
procreation, and their effectiveness is modified, giving rise to breeding with novel traits. After
that, these posterities experience a similar genetic expression cycle. This iterative procedure
repeats several times until it reaches a workable solution.
 of
ro
-p
re
lP

Fig. 6. GEP Working Mechanism

To effectively use GEP for problem-solving, it is necessary to define five key components: the
a

function set, terminal set (comprising variables for input and constants), suitability function,
rn

GEP control parameters (such as population density and crossover), and termination rules. The
initial population is dependent on the function and terminal parameters. A typical chromosome
u

has a head and tail, but Heads have functions or terminal symbols, whereas tails have just
Jo

terminal symbols. ETs of varying sizes show chromosomes whose language understanding
needs gene language sequence or organization. Functions (+, −, /, *) connect multiple genes on
a chromosome to form complex ETs. The fitness function assesses population performance by
and maximum generation number. Iteration termination or overfitting selects increasing ETs.
When the population reaches the maximum number of generations, the software pauses. If not,
it continues via reproduction, crossover, and mutation to create new populations until the
optimum standards are met or the computer finds an ideal solution.
3.1.3. Multi-expression programming (MEP)

Multi-expression programming (MEP), a distinct evolutionary method, was recently

introduced by Oltean [103]. The way that the GEP and MEP algorithms handle linearity is
where they diverge most. In MEP, each chromosome has a fixed number of expressions, known
as genes, consisting of variable-length strings. This factor considerably boosts the probability
of issue-solving. MEP treats individuals as variable-sized units and writes the computing result
as a linear form of string commands that combine terminal variables and mathematical
operators. Fig. 7 illustrates the primary stages of the MEP algorithm, and Fig. 8 demonstrates
the MEP algorithm architecture.

of
ro
-p
re
a lP
rn

Fig. 7. MEP Algorithm Flowchart

u

The MEP evaluation process involves creating a random chromosome population, using a
Jo

binary competition process to select two parents and rearrange them according to crossover
likelihood. After that, the two offspring born from the combination of the chosen parents are
altered to replace the weaker member with the strongest. The iterative process continues until
convergence is attained [103], [104]. The genes in MEP encode functions or terminals whose
indices are always lower than the chromosomal location of the gene. Each gene in the
population is decoded by classical MEPs, which then assign a fitness value depending on how
well the gene resolves the problem during selection. Next, the most optimal chromosomal
fitness assignment solution is chosen.
 Fig. 8. MEP Algorithm Architecture

of
3.2. Research strategy
The workflow of the strategy employed in this study includes data collecting, preprocessing,

ro
building ML models, and choosing and interpreting those models. The data collection phase
concentrates on variable selection that affects the CS of 3DP-FRC. These variables include raw
-p
ingredients, mix proportions, fiber characteristics, and loading directions. Data preparation
includes handling missing values, eliminating duplicate values, and standardizing the data to
re
provide consistent feature scales and reduce model bias. To simplify the evaluation of the
model, the dataset partitions into a training set, which constitutes 70% of the data, and a testing
lP

set, which makes up the remaining 30%. The DT model and SHAP are implemented using
Python's sci-kit-learn package (version 0.23.2) for implementation. Pandas (version 1.1.3) and
a

NumPy (version 1.19.2) are relevant libraries. Then, SHAP analysis interprets these ML
models, clarifying each input variable's relative impact on the output objectives. In addition,
rn

GeneXpro tool 5.0 was used for GEP modeling, while MEPX 2023.3.18 was employed for
u

MEP modeling.
Jo

3.3.1. Model development

Identifying applicable variables is a critical step in creating a trustworthy prediction model.
Multiple trials and consider earlier studies' recommendations while determining the GEP
approach's fitting parameters. In building the GEP model, the three governing parameters are
numerical constants, genetic operators, and general model parameters. General parameters
include the chromosomes, genes, the set of functions, the size of the head, and the linking
operator. This study used GEP modeling with a population size of 1000. The population size
has a direct influence on how long the process takes. The gene count determines the complexity
of each term, but the head size regulates the structure of the model by affecting the number of
sub-ETs. The current investigation adjusts the GEP model head size to 10 and gene number to
4. Table 4 comprehensively explains the GEP configuration settings, including genetic
operators. GEP-based method implementation uses GeneXpro Tool 5.0.

Table 4. Setting parameters of the GEP model

Parameter Settings Parameter Settings
General Numerical Constants
Genes 4 Constant per gene 10
Head Size 10 Upper bound 10
Chromosomes 500 Data type Floating-Point
Linking function Multiplication Lower Bound -10

of
Set of functions +, -, *, /, exp, x2, x3, 3Rt

ro
Genetic Operations
Mutation 0.0013
IS transportation rate
Inverse rate
-p
0.0054
0.0054
re
Permutation 0.0054
Random cloning 0.0010
lP

Gene transportation rate 0.0027

RIS transportation rate 0.0054
a

Recombination rate 0.0027

rn

Dc Mutation 0.0020
RNC mutation 0.0020
u
Jo

Carefully examining different MEP setup settings is necessary to develop a robust model in the
context of MEP modeling. These produced settings are recommended by several trial runs since
the number of programs developed is dependent on a critical population size parameter [105].
A larger population size might result in a complicated yet accurate model even if convergence
takes more time. However, overfitting the model may occur when population size exceeds a
particular range [91].Table 5 lists the selected setup variables for the current development of
the MEP model. Basic operations in the model include *, /, -, and +. Before the procedure
finishes, the generation quantity determines the model's intended level of accuracy. Several
generations of algorithms have helped produce a simulation model with few flaws. The current
approach trailed various combinations of parameters to find the best model and developed the
one with the lowest error values. A central problem in ML modeling is overfitting, which is the
situation when the model performs well on the original data but poorly on newly acquired data.
It is advisable to assess the model's performance on untested data to solve this problem and
understand its generalization abilities [106], [107]. To address the problem of overfitting, two
groups divide the data, with 70% for training and 30% for testing. An independent testing set
not utilized in the model's creation evaluates the algorithm's performance. The MEPX version
2023.4.14. software program implemented the MEP model development procedure.

Table 5. Setup for MEP model development

of
Parameter Setting
No of subpopulations 100

ro
Subpopulation size 150
Tournament size 2
Code length
Crossover probability
50
0.9
-p
re
Functions probability 0.5
Variables probability 0.5
lP

Functions +,-,*,\, Power, Sqrt, Exp, Sin, Cos, Tan

Mutation probability 0.01
a

3.3. Model evaluation indicators

rn

Eq.(2)-(6) represent several statistical measures, including R, MAE, RMSE, a10, and a20
indexes used to evaluate ML models' precision and effectiveness. These statistical metrics are
u

frequently used in machine learning to measure error and accuracy [108]. Increased model
Jo

efficiency is indicated by lower RMSE and MAE scores.

∑n (ei-mi)2 (2)
RMSE =√ i=1
n
∑ni=1 |ei-mi|
MAE = (3)
n
∑ni=1 ̅ )(mi-mi
(ei-ei ̅̅̅)
R=
(4)
̅ )2 ∑ni=1 (mi-mi
√∑ni=1 (ei-ei ̅̅̅)2

m10
a10 index = (5)
n
 m20
a20 index = (6)
n
In this case, "n" denotes the number of data points, "ei" denotes the original value, and "mi" is
the predicted value. Furthermore, the e𝑖
̅ and mi represent the average of the actual and predicted
results, respectively. A higher correlation coefficient (R) value shows the correctness of the
mode, which calculates the relationship between real and model data [109]. An R-value above
0.8 shows a significant association, which indicates a high relationship between the estimated
and experimental values [110], [111]. The R by itself is insufficient to evaluate the model's
overall effectiveness. Therefore, MAE and RMSE were also used for determining the model
efficacy. It is necessary to decrease estimates with significant errors when the RMSE value is

of
more prominent. In contrast, MAE regularly produces lower results than RMSE and gives

ro
substantial mistakes comparatively less weight.

3.4. Interpretability of the models

-p
SHAP is a game theory-based method for elucidating the operation of ML models. SHAP
provides valuable insights into how ML models operate. By quantifying the impact of each
re
feature in a model-agnostic way, SHAP enhances interpretability and trust in complex models
like tree-based models or neural networks. Its versatility extends to various tasks, including
lP

feature engineering, model selection, and understanding model biases, making it a valuable
tool in both research and applied settings. Tree SHAP, Kernel SHAP, and Deep SHAP are
a

techniques that can calculate SHAP values. For individual estimations, these values guarantee
rn

consistent interpretations of the influence of relationships. The dual interpretability of SHAP

values at the local and global levels is a significant benefit. SHAP has the ability to distinguish
u

between the positive and negative contributions of each input attribute, in contrast to other ML
Jo

feature significance measures. Moreover, every data point may be linked to its own SHAP
value, allowing local and global interpretation of the model. Several experts in the scientific
community have presented more detailed explanations of SHAP [112], [113]. The SHAP
mapping function is illustrated in Fig. 9.
 of
Fig. 9. SHAP attributes
In addition to SHAP, PDP, and ICE methods are also employed for enhanced model

ro
interpretation. PDP plots display the average impact of a single feature on the model's output
while controlling for other variables. It reveals whether the relationship is linear, nonlinear, or
-p
exhibits interactions with other variables. Similarly, ICE plots show predicted outcomes for
individual data points as the feature of interest changes, holding other features constant. ICE
re
plots allow examination of how the model's predictions vary across different instances,
capturing any heterogeneity or nonlinearity in the relationship between the feature and the
lP

predicted outcomes.

4. Results and discussion

a

4.1. Mathematical formulation for CS of 3DP-FRC

rn

The GeneXpro tool generated Expression Trees (ETs) for the suggested GEP model, as shown
u

in Fig. 10. Next, an empirical expression is established to estimate the strength of the 3DP-
FRC by decoding GEP-generated ETs. This expression includes mathematical functions like
Jo

exp, x2, x3, and 3Rt and fundamental arithmetic operations like addition, subtraction, and
multiplication. In GEP modeling, multiplication served as the linking function.
 of
ro
-p
re
a lP
u rn

Fig. 10. Expression Trees of the GEP model

Jo

The results of the GEP approach using CS as an ET are shown in Fig. 10. Within the ETs,
variables are represented by terminals, while mathematical operators (such as −,+, ÷, and x)
are represented by nodes. The ETs are translated by the system. The empirical expression for
predicting the strength of 3DP-FRC is developed by decoding the ETs of the GEP model to
CC+ code, which is then simplified. Furthermore, the proposed expression is explained by 10,
500, and 4 genes, respectively, in relation to head size and chromosome count. A simplified
formulation, as given in Eq. (7), is intended to provide a reliable prediction of the strength of
3DP-FRC.
 CS = X+Y+Z (7)

Where;

(W/B+(2.49583305077535+((Lf+(CA*HMPC))
X= (8)
((2.49583305077535*Df)-(Lf/Df- Df))
3
Y = √CA (9)

Z = ((∛(((((HMPC*LD)+ SF)+GS)+GS))-W/B)+FT) (7)

of
4.2. Performance of models

ro
4.2.1. Regression slope analysis

The accuracy of the suggested model is evaluated by assessing the slope of the regression line
-p
obtained from experimental data indicated on the x-axis versus prediction data values on the y-
axis. This approach is widely used by researchers [114], [115], [116], [117], [118] to assess
re
how accurate ML models are. The GEP method in Fig. 11 shows regression slope (RS) values
of 0.993 and 0.989 for the training and testing sets, respectively. Similarly, the MEP model
lP

exhibited RS values of 0.946 and 1.001 for the training and testing subsets, respectively. The
DT model showed an RS value of 0.983 for training and 1.002 for testing. All models showed
a

regression slopes larger than 0.8 and closer to 1 for training and testing, suggesting a high
rn

correlation between the model's estimated and actual values. Notably, the GEP model showed
RS values comparatively closer to 1 for both subsets, indicating its precise estimation efficacy
u

in estimating CS of 3DP-FRC. In addition, the fitting lines for training and testing subsets of
Jo

the GEP and DT models are closely aligned with the ideal fit lines, indicating greater
performance of these models compared to the MEP model.
 of
ro
-p
re
a lP
u rn
Jo

Fig. 11. Regression slope analysis of the models

4.2.2. Error assessment

Fig. 12 depicts the error evaluation of the GEP model, presenting the error histograms and error
representation plots of the established models. Notably, the GEP model predictions are closely
aligned with the actual data. The GEP model exhibited 97.85% of the prediction’s records
within the error range of ±2.5 MPa. Similarly, the MEP model showed 68.23% of the
predictions within the error range of ±10 MPa, as given in Fig. 13. The error histogram for the
DT model shows that 88.29% of the records fall in the range of -5 and 5 MPa. Overall, the GEP
model showed the lowest error while the MEP model exhibited a higher error, indicating the
potential of the GEP model to estimate the strength of 3DP-FRC with minimal error.
 200
Experimental Training Testing
180 GEP predicted 250
Error 222
160
140 200
120

Frequency
CS (MPa)

100 150
80
60 100
40
20 50
26
0 11 15
4 4 8 4
−20 0

of
0 50 100 150 200 250 300 −10.0 −7.5 −5.0 −2.5 0.0 2.5 5.0 7.5 10.0
Dataset Error range (GEP model)

ro
Fig. 12. Error analysis of the GEP model

200
Experimental
MEP predicted
Training Testing
-p 100
101
re
Error
150
80
CS (MPa)

100
Frequency
lP

60 54
49
50
40 35
a

0 18 18
20
rn

−50 0
0 50 100 150 200 250 300 −15 −10 −5 0 5 10 15
Dataset Erro range (MEP model)
u
Jo

Fig. 13. Error analysis of the MEP model

200
Experimental Training Testing 160 152
DT predicted
Error
140
150
120
Frequency
CS (MPa)

100 100 97

80
50
60

40
0
20 19 15
8 5
−50 0
0 50 100 150 200 250 300 −15 −10 −5 0 5 10 15
Dataset Error range (DT model)

Fig. 14. Error analysis of the DT model

4.2.3. Statistical metrics assessment

Multiple performance metrics were employed to determine the generalizability and accuracy
of the established models. The summary of these utilized is provided in Table 6. The analysis
shows that the GEP model consistently outperforms the other models, like the MEP and DT
models, in estimating the output. GEP achieves high R-values of 0.995 for the testing set and
0.996 for training sets, indicating its strong predictive power. It also has the lowest MAE, with
values of 2.951 for training and 2.673 for testing. This means GEP is the most accurate in
predicting compressive strength. Additionally, GEP performs better in terms of a10 index and
a20 index. Overall, the GEP model demonstrated higher values for R, a10 index, and a20 index,

of
while also showcasing the lowest values for MAE and RMSE in both training and testing sets.
In summary, the statistical analysis collectively underscores the unparalleled efficacy of the

ro
GEP model in accurately estimating the CS of 3DP-FRC, thereby highlighting its immense
potential for practical applications in the field. Fig. 15 presents spider plots depicting the
-p
graphical representation of statistical indicator scores.
re
Table 6. Performance assessment of the developed models
lP

Model Subset RMSE MAE R a10 Index a20 Index

Training 2.951 1.163 0.995 0.837 0.876
GEP
Testing 2.673 1.029 0.996 0.778 0.833
a

Training 9.319 6.798 0.951 0.837 0.703

MEP
rn

Testing 6.671 4.832 0.973 0.444 0.733

Training 4.755 2.606 0.987 0.837 0.847
u

DT
Testing 4.071 2.513 0.990 0.622 0.833
Jo
 RMSE Training RMSE
Testing
7 10
GEP
6 GEP
MEP 8 MEP
5 DT
DT
6
a20-index 4 6 a20-index
5 MAE 8 MAE
0.85 4 6
3 3 0.90 4 4
0.80 2 0.85 2
1 0.80
0.75 0.75
0.70

0.5 0.98 0.96

0.6 0.5
0.99 0.98
0.7 1.0

of
0.8 1.00 1.5 1.00
0.9
a10-index R a10-index R

ro
Fig. 15. Spider Diagrams showing Statistical Indicator Scores

4.3. Comparison of the suggested models

-p
The accuracy of the suggested approaches is compared to assess their relative prediction
re
performance in predicting the output. In the training, the GEP model exhibits substantial
superiority with MAE scores that are 82.09% and 44.62% less than those of the MEP and DT
lP

approaches, respectively. In the testing phase, the GEP model showed MAE scores that were
78.70% and 59.05% less than those of the MEP and DT methods, respectively. This significant
reduction in error further solidifies the GEP model's efficacy and reliability in accurately
a

predicting outcomes even on unseen data.

rn

To enhance the comparison of the established models' performance, Fig. 16 displays Taylor
u

diagrams, commonly used to assess the effectiveness of predictive models. When evaluating
the effectiveness of ML models, researchers frequently use the Taylor diagram for model
Jo

assessment [119]. Key performance measures from various datasets, including the centered
RMS error, R, and SD, constitute the foundation of this evaluation. The actual experimental
data represented by the red color symbol serves as a reference point. The best model is the one
that shows the closest distance to this reference point. The GEP symbol is positioned closer to
the benchmark (marked in red) in the graph, followed by DT and then MEP models in both
testing and training sets. Based on the Taylor diagram analysis, the models can be ranked as
follows: GEP > DT > MEP. This ranking indicates that the GEP model outperforms both the
DT and MEP models in terms of their overall performance and alignment with the observed
data.
 Testing set
Training set
0 0.2 0 0.2 CS (Ref)
42.0 0.4 CS (Ref.) 0.4
41.0

Co
Co
GEP
0.6 0.6

rr
GEP

rr
MEP

ela
ela
MEP
DT

tio
tio

Standard Deviation
Standard Deviation

41.5 DT 40.8

n
Co

Co
0.8 0.8

eff

eff
ice

ice
40.6

nt

nt
0.9 0.9
41.0
40.4
0.95 0.95

40.5 0.99 40.2

0.99

1 40.0 1
40.5

41.0

41.5

42.0

40.0

40.2

40.4

40.6

40.8

41.0
Standard Deviation Standard Deviation

of
Fig. 16. Taylor Diagrams Representing the Model's Comparison

ro
4.4. Comparison with literature models
Table 7 compares the current models to those developed in the literature for the CS of 3DP
concrete. All the researchers listed in Table 7 used the same dataset except Izadgishasb et al.
[121] who utilized a different but still comparison dataset with less data points as compared to
-p
re
the other three studies.Uddin et al. [120] predicted the strength of 3DP-FRC using different
types of ML approaches. The CatBoost method performed better, showing an exceptional R of
lP

0.99 and less errors (MAE=3.14, RMSE=5.17). Similarly, Izadgoshasb et al. [121] used
Artificial Neural Networks (ANN) and the Multi-Objective Grasshopper Optimization
a

Algorithm (MOGOA) to CS of 3D-printed mortar. The ANN-MOGOA model performed

rn

exceptionally well, with an R-value of 0.98. The MAE and RMSE values were 2.58 and 4.49,
respectively. Ghasemi and Naser [122] employed the XGBoost, RF, and MLR models to
u

forecast the strength of 3DP-FRC. Furthermore, Alyami et al. [123] used four different
Jo

algorithms and reported that gradient boosting (GB) provided higher accuracy.

In comparison to the excellent-performing model reported in the previous studies, namely

LightGBM presented by Uddin et al. [120], the GEP model established in present case
showcased superior efficacy in estimating the strength of 3DP-FRC. For example, the MAE
score of the GEP model is 58.84% lower than that of the LightGBM model, while the RMSE
is 45.49% lower than that of the LightGBM model. This finding suggests that GEP method
developed in our research outperforms even the state-of-the-art model identified in prior
studies, underlining its effectiveness and potential for accurate CS prediction in 3DP-FRC
applications. Furthermore, the developed DT and MEP models showed comparable accuracy
to the literature models.
Furthermore, in this study, we developed a simplified GEP-based equation tailored specifically
for estimating the CS of 3DP-FRC. This streamlined equation offers a user-friendly solution,
allowing for straightforward implementation and accurate CS predictions. In addition, while
SHAP was predominantly employed in the literature, the interpretation of machine learning
(ML) models was further enriched by incorporating additional methods such as Partial
Dependence Plots (PDP) and Individual Conditional Expectation (ICE). This broadened
approach allowed for a comprehensive exploration of model predictions and a deeper
understanding of the factors influencing the estimated compressive strength of 3DP-FRC.

of
Table 7. Comparison of the established models with literature
models

ro
Model R RMSE MAE Reference
RF 0.96 10.89 8.05
SVM
XG Boost
0.96
0.99
-p
11.97
5.50
7.55
3.15
re
[120]
LightBGM 0.99 4.94 2.50
NGBoost 0.98 6.33 4.55
lP

CatBoost 0.99 5.17 3.14

SVR 0.84 18.72 10.25
a

DT 0.99 6.60 4.64

[123]
0.99 7.21 3.90
rn

GB
RF 0.99 7.13 3.99
u

MOGOA and ANN 0.98 4.49 2.58 [121]

0.91 14.60 8.37
Jo

RF
XGBoost 0.92 8.70 6.00 [122]
MLR 0.65 20.14 16.14
GEP 0.996 2.673 1.029
MEP 0.973 6.671 4.832
DT 0.990 4.071 2.513

4.5. SHAP analysis

The model interpretation using the SHAP approach is the main topic of this section. DT
approach prediction results are considered for SHAP analysis as DT showed the best accuracy
in predicting the CS of 3DP-FRC.
Fig. 17 illustrates the critical portion that every input feature plays in determining the model's
output (CS). Water content (W) has a notable influence on the CS of 3DP-FRC, as seen by its
considerable SHAP value of +17.16, just above W/B, which is +15.96Similarly, with their
respective SHAP values of +5.39 and +4.11, SP and SF show significant contributions,
highlighting their crucial roles in predicting the strength of 3DP-FRC. Additionally,
characteristics like LD, VF, S, GA, and CA show moderate-level SHAP values, highlighting
their impact on the CS of 3DP-FRC. Although the Fig. 17 feature significance plot offers a
broad overview of the relevance of the input characteristics, it does not clarify whether each
feature has a positive or negative impact on the output.

of
ro
-p
re
a lP
u rn
Jo

Fig. 17. SHAP values plot.

A SHAP summary plot of each feature's effect on compressive strength is shown graphically
in Fig. 18. The y-axis represents features in descending order of importance, while the x-axis
shows the associated SHAP values. Each dot on this plot represents a distinct data sample. The
color of the dots corresponds to the signal's intensity, with blue indicating lower magnitudes
and red indicating higher ones. Inspection reveals that the W red dots are on the left (negative),
negatively impacting the 3DP-FRC's CS, which indicates that compressive strength reduces
with increasing water content, which aligns with the findings of Fig. 2, where W shows a
negative correlation of -0.82 with CS. On the other hand, LD, CA, SP, and Df positively impact
the compressive strength. These observations highlight several characteristics' complex and
variable influences on the total compressive strength.

of
ro
-p
re
a lP
u rn
Jo

Fig. 18. SHAP summary plot

4.6. ICE and PDP analysis

In addition to the SHAP approach, Fig. 19 and Fig. 20 show the results of PDP and ICE
analyses for the constructed model. PDP and ICE analysis clarify model predictions. PDP
shows the mean influence of each feature on the model's predictions, revealing how specific
traits affect the outcome while adding other factors. ICE analysis provides a deeper
understanding of predictions for each dataset instance, enabling the identification of variation
in prediction behavior across distinct data points. These strategies increase interpretation and
provide a complete grasp of the model's decision-making process, enhancing confidence about
its forecasts.

The CS exhibits a pronounced decline when the water-to-binder ratio (W/B) increases above
0.24. Furthermore, the CS continues to fall gradually, indicating that neither water nor binder
has a substantial individual impact, but rather, their combined influence is responsible for this
trend. The SP curve generally reveals little fluctuations like the SF curve, except for a tiny
upward shift at a content value of 2.4 (Kg/m3), which aligns with the findings of Arunothayan
et al. [33]. Chu et al. [68] found that the influence of SF on CS is minimal. Zhou et al. [85]
suggested using Hybrid materials, such as carbon-embedded fibers, in 3DFRC to enhance its

of
properties. Weng et al. [124] discovered that raising the SP dose from 0.15 to 0.30 w%
increases 3DP-FRC lamination adhesive strength from 2.7 to 5.7 MPa. Although its primary

ro
objective is flowability, the SP also improves 3DP-FRC compressive strength [125].
Additionally, in the case of LD, the CS steadily rises to a value of 10 and beyond, exhibiting
-p
an inverse pattern and a slope of almost half the initial rise. Arunothayan et al. [48] studied the
fiber orientation of 3DP-FRC under various loading orientations and found that the maximum
re
strength showed up in the printing direction This finding supports the concept of dead weight
self-compaction phenomena, previously explored by Delvasto et al. [126] The reason for this
lP

could be a stronger interlayer connection caused by pressure-induced layer extrusion,

especially when the printing height is lower than the nozzle size, along with the intricate
a

interaction between loading orientations and steel fibers. However, the CS for VF exhibits early
rn

fluctuations up to 0.08 % and thereafter remains constant with little impact from VF. It is
important to observe that the CS reduces as the VF increases up to 0.005 %. After that, it
u

doubled its previous value and remained stable. CS is not affected by S, GS, CA, AR, HPMC,
Jo

and FT content. Furthermore, OPC has a dynamic impact on CS as it steadily rises to 450
(Kg/m3), then instantly levels off at 650 (Kg/m3), and then spikes, reaching a peak of 980
(Kg/m3) before stabilizing. When approaching DF, a lesser number of up to 40 (µm) has an
impact, while a more significant amount does not affect CS. While LF exhibits a negligible
impact, the inclusion of FA content lacks any considerable influence on CS.
 of
ro
-p
re
a lP
rn
u
Jo

Fig. 19. PDP analysis

Jo
urn
alP
re
-p
ro
of
 Fig. 20. ICE plots

of
5. Conclusion
This study aimed to develop robust and accurate prediction models for estimating the

ro
compressive strength of fiber-reinforced concrete that has been 3D printed. Three different
machine learning models, MEP, DT, and GEP, collectively developed these models. Sixteen
-p
input variables were considered were considered for development. Various statistical measures
assessed the model's generalizability and accuracy, such as the R, MAE, RMSE, a10 Index, and
re
a20 Index. In addition, the SHAP, ICE, and PDP approaches are employed to interpret the
predictions of the model, which provides valuable insights. The study's main conclusions are
lP

herein:

1. With R-values of 0.93, 0.93, and 0.98 for training and flat 0.97 for testing, respectively,
a

the predictive models GEP, MEP, and DT demonstrated impressive accuracy,

rn

demonstrating their effectiveness in forecasting the strength of 3DP-FRC.

2. The GEP technique generated an empirical formulation that presents a simple
u

mathematical model for estimating CS in 3DP-FRC. This formulation enhances the

Jo

prediction model's practical usefulness.

3. W, SA, DF, CA, and LD were crucial input factors that significantly impacted the
prediction of compressive strength in 3DP-FRC by the SHAP analysis. These results
point to essential variables to consider while maximizing the mix design for 3DP-FRC.
4. By utilizing ICE and PDP interpretations, the research explores the complex
interrelationships among various input factors and identifies their unique effects on the
compressive strength of 3DP-FRC. This comprehensive interpretability offers critical
information for determining the ideal mix percentage in 3DP-FRC, which improves
performance in real-world applications.

This work has significantly advanced predictive modeling for the compressive strength of 3D-
FRC. The practical value of ML models' better accuracy, the identification of crucial input
variables, and the interpretability offered by the different explainability approaches are
noteworthy for researchers and engineers. However, hybrid machine learning models present
an exciting potential for future research and development, perhaps leading to even more robust
and reliable prediction models. Moreover, controlled experiments must be carried out in future
research projects to strengthen the validity and resilience of these models. Since the current
study used a variety of literature sources with different experimental settings, it becomes
imperative to gather data from a single, reliable source in a consistent context.

References:

of
[1] A. Perrot, D. Rangeard, and A. Pierre, “Structural built-up of cement-based materials
used for 3D-printing extrusion techniques,” Mater Struct, vol. 49, no. 4, pp. 1213–1220,

ro
Jan. 2016, doi: 10.1617/s11527-015-0571-0.
[2] B. Han, L. Zhang, and J. Ou, Smart and multifunctional concrete toward sustainable

[3]
-p
infrastructures. 2017. doi: 10.1007/978-981-10-4349-9.
P. Hosseini, R. Hosseinpourpia, A. Pajum, M. M. Khodavirdi, H. Izadi, and A. Vaezi,
re
“Effect of nano-particles and aminosilane interaction on the performances of cement-
based composites: An experimental study,” Constr Build Mater, vol. 66, pp. 113–124,
lP

Jan. 2014, doi: 10.1016/j.conbuildmat.2014.05.047.

[4] B. Panda, S. C. Paul, and M. J. Tan, “Anisotropic mechanical performance of 3D printed
a

fiber reinforced sustainable construction material,” Mater Lett, vol. 209, pp. 146–149,
rn

Jan. 2017, doi: 10.1016/j.matlet.2017.07.123.

[5] R. A. Buswell, W. R. Leal de Silva, S. Z. Jones, and J. Dirrenberger, “3D printing using
u

concrete extrusion: A roadmap for research,” Cem Concr Res, vol. 112, pp. 37–49, Oct.
Jo

2018, doi: 10.1016/J.CEMCONRES.2018.05.006.

[6] S. Hou, Z. Duan, J. Xiao, and J. Ye, “A review of 3D printed concrete: Performance
requirements, testing measurements and mix design,” Constr Build Mater, vol. 273, p.
121745, Jan. 2021, doi: 10.1016/j.conbuildmat.2020.121745.
[7] B. Zareiyan and B. Khoshnevis, “Interlayer adhesion and strength of structures in
Contour Crafting - Effects of aggregate size, extrusion rate, and layer thickness,” Autom
Constr, vol. 81, 2017, doi: 10.1016/j.autcon.2017.06.013.
[8] J. G. Sanjayan, B. Nematollahi, M. Xia, and T. Marchment, “Effect of surface moisture
on inter-layer strength of 3D printed concrete,” Constr Build Mater, vol. 172, 2018, doi:
10.1016/j.conbuildmat.2018.03.232.
[9] P. Feng, X. Meng, J.-F. Chen, and L. Ye, “Mechanical properties of structures 3D printed
with cementitious powders,” Constr Build Mater, vol. 93, pp. 486–497, Jan. 2015, doi:
10.1016/j.conbuildmat.2015.05.132.
[10] V. S. F. Ronald Rael, “Printing Architecture: Innovative Recipes for 3D Printing,” J
Chem Inf Model, vol. 53, no. 9, 2018.
[11] D. Lowke, E. Dini, A. Perrot, D. Weger, C. Gehlen, and B. Dillenburger, “Particle-bed
3D printing in concrete construction – Possibilities and challenges,” Cem Concr Res,
vol. 112, pp. 50–65, Jan. 2018, doi: 10.1016/j.cemconres.2018.05.018.
[12] S. Lim, R. A. Buswell, T. T. Le, S. A. Austin, A. G. F. Gibb, and T. Thorpe,
“Developments in construction-scale additive manufacturing processes,” Autom Constr,

of
vol. 21, pp. 262–268, Jan. 2012, doi: 10.1016/j.autcon.2011.06.010.

ro
[13] M. Gomaa, W. Jabi, V. Soebarto, and Y. M. Xie, “Digital manufacturing for earth
construction: A critical review,” J Clean Prod, vol. 338, p. 130630, Jan. 2022, doi:

[14]
10.1016/j.jclepro.2022.130630.
-p
M. P. A. M. Marijnissen and A. van der Zee, “3D Concrete Printing in Architecture A
re
research on the potential benefits of 3D Concrete Printing in Architecture,” in
Proceedings of the International Conference on Education and Research in Computer
lP

Aided Architectural Design in Europe, 2017. doi: 10.52842/conf.ecaade.2017.2.299.

[15] S. Al-Qutaifi, A. Nazari, and A. Bagheri, “Mechanical properties of layered geopolymer
structures applicable in concrete 3D-printing,” Constr Build Mater, vol. 176, pp. 690–
a

699, Jan. 2018, doi: 10.1016/j.conbuildmat.2018.04.195.

rn

[16] T. T. Le et al., “Hardened properties of high-performance printing concrete,” Cem Concr

Res, vol. 42, no. 3, pp. 558–566, Jan. 2012, doi: 10.1016/j.cemconres.2011.12.003.
u

[17] Y. W. D. Tay, G. H. A. Ting, Y. Qian, B. Panda, L. He, and M. J. Tan, “Time gap effect
Jo

on bond strength of 3D-printed concrete,” Virtual Phys Prototyp, vol. 14, no. 1, pp. 104–
113, Jan. 2019, doi: 10.1080/17452759.2018.1500420.
[18] B. Panda, S. C. Paul, N. A. N. Mohamed, Y. W. D. Tay, and M. J. Tan, “Measurement of
tensile bond strength of 3D printed geopolymer mortar,” Measurement, vol. 113, pp.
108–116, Jan. 2018, doi: 10.1016/j.measurement.2017.08.051.
[19] B. Panda, S. C. Paul, L. J. Hui, Y. W. D. Tay, and M. J. Tan, “Additive manufacturing of
geopolymer for sustainable built environment,” J Clean Prod, vol. 167, 2017, doi:
10.1016/j.jclepro.2017.08.165.
[20] J. C. Lin, X. Wu, W. Yang, R. X. Zhao, and L. G. Qiao, “The influence of fine aggregates
on the 3D printing performance,” IOP Conf Ser Mater Sci Eng, vol. 292, p. 12079, Jan.
2018, doi: 10.1088/1757-899X/292/1/012079.
[21] M. Hambach and D. Volkmer, “Properties of 3D-printed fiber-reinforced Portland
cement paste,” Cem Concr Compos, vol. 79, pp. 62–70, Jan. 2017, doi:
10.1016/j.cemconcomp.2017.02.001.
[22] D. G. Soltan and V. C. Li, “A self-reinforced cementitious composite for building-scale
3D printing,” Cem Concr Compos, vol. 90, pp. 1–13, Jan. 2018, doi:
10.1016/j.cemconcomp.2018.03.017.
[23] A. Inozemtcev, E. Korolev, and D. T. Qui, “Study of mineral additives for cement
materials for 3D-printing in construction,” IOP Conf Ser Mater Sci Eng, vol. 365, p.
32009, Jan. 2018, doi: 10.1088/1757-899X/365/3/032009.
[24] G. Ma and L. Wang, “A critical review of preparation design and workability

of
measurement of concrete material for largescale 3D printing,” Frontiers of Structural

ro
and Civil Engineering, vol. 12, no. 3, pp. 382–400, Jan. 2018, doi: 10.1007/s11709-017-
0430-x.

-p
[25] R. A. Buswell, W. R. L. de Silva, S. Z. Jones, and J. Dirrenberger, “3D printing using
concrete extrusion: A roadmap for research,” Cem Concr Res, vol. 112, pp. 37–49, Jan.
re
2018, doi: 10.1016/j.cemconres.2018.05.006.
[26] T. T. Le, S. A. Austin, S. Lim, R. A. Buswell, A. G. F. Gibb, and T. Thorpe, “Mix design
lP

and fresh properties for high-performance printing concrete,” Materials and

Structures/Materiaux et Constructions, vol. 45, no. 8, 2012, doi: 10.1617/s11527-012-
9828-z.
a

[27] G. Ma, Z. Li, and L. Wang, “Printable properties of cementitious material containing
rn

copper tailings for extrusion based 3D printing,” Constr Build Mater, vol. 162, 2018,
doi: 10.1016/j.conbuildmat.2017.12.051.
u

[28] A. Kazemian, X. Yuan, E. Cochran, and B. Khoshnevis, “Cementitious materials for

Jo

construction-scale 3D printing: Laboratory testing of fresh printing mixture,” Constr

Build Mater, vol. 145, 2017, doi: 10.1016/j.conbuildmat.2017.04.015.
[29] M. Xia and J. Sanjayan, “Method of formulating geopolymer for 3D printing for
construction applications,” Mater Des, vol. 110, pp. 382–390, Jan. 2016, doi:
10.1016/j.matdes.2016.07.136.
[30] V. Mechtcherine, J. Grafe, V. N. Nerella, E. Spaniol, M. Hertel, and U. Füssel, “3D-
printed steel reinforcement for digital concrete construction – Manufacture, mechanical
properties and bond behaviour,” Constr Build Mater, vol. 179, pp. 125–137, Jan. 2018,
doi: 10.1016/j.conbuildmat.2018.05.202.
[31] J. Zhang, J. Wang, S. Dong, X. Yu, and B. Han, “A review of the current progress and
application of 3D printed concrete,” Compos Part A Appl Sci Manuf, vol. 125, p. 105533,
Jan. 2019, doi: 10.1016/j.compositesa.2019.105533.
[32] T. A. M. Salet, Z. Y. Ahmed, F. P. Bos, and H. L. M. Laagland, “Design of a 3D printed
concrete bridge by testing*,” Virtual Phys Prototyp, vol. 13, no. 3, 2018, doi:
10.1080/17452759.2018.1476064.
[33] A. R. Arunothayan, B. Nematollahi, R. Ranade, K. H. Khayat, and J. G. Sanjayan,
“Digital fabrication of eco-friendly ultra-high performance fiber-reinforced concrete,”
Cem Concr Compos, vol. 125, p. 104281, Jan. 2022, doi:
10.1016/j.cemconcomp.2021.104281.

of
[34] D. Avrutis, A. Nazari, and J. G. Sanjayan, “Chapter 19 - Industrial Adoption of 3D

ro
Concrete Printing in the Australian Market: Potentials and Challenges,” in 3D Concrete
Printing Technology, 2019.

concrete: applications, performance,

-p
[35] A. Siddika, M. A. Al Mamun, W. Ferdous, A. K. Saha, and R. Alyousef, “3D-printed
and challenges,” 2020. doi:
re
10.1080/21650373.2019.1705199.
[36] T. Marchment and J. Sanjayan, “Mesh reinforcing method for 3D Concrete Printing,”
lP

Autom Constr, vol. 109, p. 102992, Jan. 2020, doi: 10.1016/j.autcon.2019.102992.

[37] Z. Li, L. Wang, and G. Ma, “Mechanical improvement of continuous steel microcable
reinforced geopolymer composites for 3D printing subjected to different loading
a

conditions,” Compos B Eng, vol. 187, p. 107796, Jan. 2020, doi:

rn

10.1016/j.compositesb.2020.107796.
[38] M. Liang, Z. Chang, Z. Wan, Y. Gan, E. Schlangen, and B. Šavija, “Interpretable
u

Ensemble-Machine-Learning models for predicting creep behavior of concrete,” Cem

Jo

Concr Compos, vol. 125, p. 104295, Jan. 2022, doi:

10.1016/j.cemconcomp.2021.104295.
[39] A. Singh, Q. Liu, J. Xiao, and Q. Lyu, “Mechanical and macrostructural properties of
3D printed concrete dosed with steel fibers under different loading direction,” Constr
Build Mater, vol. 323, p. 126616, Jan. 2022, doi: 10.1016/j.conbuildmat.2022.126616.
[40] F. P. Bos, E. Bosco, and T. A. M. Salet, “Ductility of 3D printed concrete reinforced with
short straight steel fibers,” Virtual Phys Prototyp, vol. 14, no. 2, 2019, doi:
10.1080/17452759.2018.1548069.
[41] G. Ma, Z. Li, L. Wang, F. Wang, and J. Sanjayan, “Mechanical anisotropy of aligned
fiber reinforced composite for extrusion-based 3D printing,” Constr Build Mater, vol.
202, pp. 770–783, Jan. 2019, doi: 10.1016/j.conbuildmat.2019.01.008.
[42] Y. Zhao, W. Meng, P. Wang, D. Qian, W. Cheng, and Z. Jia, “Research Progress of
Concrete 3D Printing Technology and Its Equipment System, Material, and Molding
Defect Control,” Journal of Engineering, vol. 2022, pp. 1–22, Jan. 2022, doi:
10.1155/2022/6882386.
[43] M. Hambach, H. Möller, T. Neumann, and D. Volkmer, “Portland cement paste with
aligned carbon fibers exhibiting exceptionally high flexural strength (> 100 MPa),”
Cem Concr Res, vol. 89, pp. 80–86, Jan. 2016, doi: 10.1016/j.cemconres.2016.08.011.
[44] C. Cakiroglu, Y. Aydın, G. Bekdaş, and Z. W. Geem, “Interpretable Predictive Modelling
of Basalt Fiber Reinforced Concrete Splitting Tensile Strength Using Ensemble Machine
Learning Methods and SHAP Approach,” Materials 2023, Vol. 16, Page 4578, vol. 16,

of
no. 13, p. 4578, Jun. 2023, doi: 10.3390/MA16134578.

ro
[45] T. Ding, J. Xiao, S. Zou, and X. Zhou, “Anisotropic behavior in bending of 3D printed
concrete reinforced with fibers,” Compos Struct, vol. 254, p. 112808, Jan. 2020, doi:

[46]
10.1016/j.compstruct.2020.112808.
-p
J. Van Der Putten, A. V. Rahul, G. De Schutter, and K. Van Tittelboom, “Development
re
of 3D Printable Cementitious Composites with the Incorporation of Polypropylene
Fibers,” Materials, vol. 14, no. 16, p. 4474, Jan. 2021, doi: 10.3390/ma14164474.
lP

[47] C. K. Y. Yu Jing and Leung, “Impact of 3D Printing Direction on Mechanical

Performance of Strain-Hardening Cementitious Composite (SHCC),” in First RILEM
International Conference on Concrete and Digital Fabrication – Digital Concrete 2018,
a

R. J. Wangler Timothy and Flatt, Ed., Cham: Springer International Publishing, 2019,
rn

pp. 255–265.
[48] A. R. Arunothayan, B. Nematollahi, R. Ranade, S. H. Bong, J. G. Sanjayan, and K. H.
u

Khayat, “Fiber orientation effects on ultra-high performance concrete formed by 3D

Jo

printing,” Cem Concr Res, vol. 143, p. 106384, Jan. 2021, doi:
10.1016/j.cemconres.2021.106384.
[49] H. Al Abadi, H. T. Thai, V. Paton-Cole, and V. I. Patel, “Elastic properties of 3D printed
fibre-reinforced structures,” Compos Struct, vol. 193, 2018, doi:
10.1016/j.compstruct.2018.03.051.
[50] G. W. Melenka, B. K. O. Cheung, J. S. Schofield, M. R. Dawson, and J. P. Carey,
“Evaluation and prediction of the tensile properties of continuous fiber-reinforced 3D
printed structures,” Compos Struct, vol. 153, 2016, doi:
10.1016/j.compstruct.2016.07.018.
[51] T. Marchment and J. Sanjayan, “Reinforcement method for 3D concrete printing using
paste-coated bar penetrations,” Autom Constr, vol. 127, 2021, doi:
10.1016/j.autcon.2021.103694.
[52] J. H. Lim, B. Panda, and Q. C. Pham, “Improving flexural characteristics of 3D printed
geopolymer composites with in-process steel cable reinforcement,” Constr Build Mater,
vol. 178, 2018, doi: 10.1016/j.conbuildmat.2018.05.010.
[53] P. Wu, J. Wang, and X. Wang, “A critical review of the use of 3-D printing in the
construction industry,” 2016. doi: 10.1016/j.autcon.2016.04.005.
[54] L. Pham, X. Lin, R. J. Gravina, and P. Tran, “Influence of pva and pp fibres at different
volume fractions on mechanical properties of 3d printed concrete,” in Lecture Notes in

of
Civil Engineering, 2021. doi: 10.1007/978-981-15-8079-6_185.

ro
[55] L. G. Li, B. F. Xiao, Z. Q. Fang, Z. Xiong, S. H. Chu, and A. K. H. Kwan, “Feasibility
of glass/basalt fiber reinforced seawater coral sand mortar for 3D printing,” Addit

-p
Manuf, vol. 37, p. 101684, Jan. 2021, doi: 10.1016/j.addma.2020.101684.
[56] V. Afroughsabet, L. Biolzi, and T. Ozbakkaloglu, “High-performance fiber-reinforced
re
concrete: a review,” 2016. doi: 10.1007/s10853-016-9917-4.
[57] C. Qi, J. Weiss, and J. Olek, “Characterization of plastic shrinkage cracking in fiber
lP

reinforced concrete using image analysis and a modified Weibull function,” Materials
and Structures/Materiaux et Constructions, vol. 36, no. 6, 2003, doi:
10.1007/bf02481064.
a

[58] P. Lura, B. Pease, G. B. Mazzotta, F. Rajabipour, and J. Weiss, “Influence of shrinkage-

rn

reducing admixtures on development of plastic shrinkage cracks,” ACI Mater J, vol.

104, no. 2, 2007, doi: 10.14359/18582.
u

[59] D. P. Bentz, “Influence of shrinkage-reducing admixtures on early-age properties of

Jo

cement pastes,” Journal of Advanced Concrete Technology, vol. 4, no. 3, 2006, doi:
10.3151/jact.4.423.
[60] R. Henkensiefken, P. Briatka, and D. Bentz, “Plastic shrinkage cracking in internally
cured mixtures made with pre-wetted lightweight aggregate,” Concrete …, vol. 32, no.
408, 2010.
[61] J. Ye, C. Cui, J. Yu, K. Yu, and F. Dong, “Effect of polyethylene fiber content on
workability and mechanical-anisotropic properties of 3D printed ultra-high ductile
concrete,” Constr Build Mater, vol. 281, 2021, doi:
10.1016/j.conbuildmat.2021.122586.
[62] A. Dulaj, T. A. M. Salet, and S. S. Lucas, “Mechanical properties and self-sensing ability
of graphene-mortar compositions with different water content for 3D printing
applications,” Mater Today Proc, vol. 70, 2022, doi: 10.1016/j.matpr.2022.09.278.
[63] C. Liu, Y. Zhang, and N. Banthia, “Unveiling pore formation and its influence on
micromechanical property and stress distribution of 3D printed foam concrete modified
with hydroxypropyl methylcellulose and silica fume,” Addit Manuf, vol. 71, 2023, doi:
10.1016/j.addma.2023.103606.
[64] J. Xiao, N. Han, L. Zhang, and S. Zou, “Mechanical and microstructural evolution of
3D printed concrete with polyethylene fiber and recycled sand at elevated temperatures,”
Constr Build Mater, vol. 293, p. 123524, Jan. 2021, doi:

of
10.1016/j.conbuildmat.2021.123524.

ro
[65] Y. Zhang and F. Aslani, “Development of fibre reinforced engineered cementitious
composite using polyvinyl alcohol fibre and activated carbon powder for 3D concrete
printing,” Constr Build Mater,
10.1016/j.conbuildmat.2021.124453.
-p
vol. 303, p. 124453, Jan. 2021, doi:
re
[66] P. Shakor, S. Nejadi, G. Paul, and J. Sanjayan, “Dimensional accuracy, flowability,
wettability, and porosity in inkjet 3DP for gypsum and cement mortar materials,” Autom
lP

Constr, vol. 110, p. 102964, Jan. 2020, doi: 10.1016/j.autcon.2019.102964.

[67] C. Liu et al., “Analysis of the mechanical performance and damage mechanism for 3D
printed concrete based on pore structure,” Constr Build Mater, vol. 314, p. 125572, Jan.
a

2022, doi: 10.1016/j.conbuildmat.2021.125572.

rn

[68] S. H. Chu, L. G. Li, and A. K. H. Kwan, “Development of extrudable high strength fiber
reinforced concrete incorporating nano calcium carbonate,” Addit Manuf, vol. 37, p.
u

101617, Jan. 2021, doi: 10.1016/j.addma.2020.101617.

Jo

[69] P. Zhang, K. Wang, J. Wang, J. Guo, S. Hu, and Y. Ling, “Mechanical properties and
prediction of fracture parameters of geopolymer/alkali-activated mortar modified with
PVA fiber and nano-SiO2,” Ceram Int, vol. 46, no. 12, pp. 20027–20037, Jan. 2020, doi:
10.1016/j.ceramint.2020.05.074.
[70] M. N. Uddin, L. Li, A. Ahmed, and K. Y. M. Almajhali, “Prediction of PVA fiber effect
in Engineered Composite cement (ECC) by Artificial neural Network (ANN),” Mater
Today Proc, vol. 65, pp. 537–542, 2022, doi: 10.1016/j.matpr.2022.03.088.
[71] C. Cakiroglu, M. Shahjalal, K. Islam, S. M. F. Mahmood, A. H. M. M. Billah, and M. L.
Nehdi, “Explainable ensemble learning data-driven modeling of mechanical properties
of fiber-reinforced rubberized recycled aggregate concrete,” Journal of Building
Engineering, vol. 76, p. 107279, Oct. 2023, doi: 10.1016/J.JOBE.2023.107279.
[72] S. Mahjoubi, R. Barhemat, P. Guo, W. Meng, and Y. Bao, “Prediction and multi-
objective optimization of mechanical, economical, and environmental properties for
strain-hardening cementitious composites (SHCC) based on automated machine
learning and metaheuristic algorithms,” J Clean Prod, vol. 329, p. 129665, Jan. 2021,
doi: 10.1016/j.jclepro.2021.129665.
[73] H. Li, J. Lin, X. Lei, and T. Wei, “Compressive strength prediction of basalt fiber
reinforced concrete via random forest algorithm,” Mater Today Commun, vol. 30, p.
103117, Jan. 2022, doi: 10.1016/j.mtcomm.2021.103117.
[74] Y. Li et al., “Compressive Strength of Steel Fiber-Reinforced Concrete Employing
Supervised Machine Learning Techniques,” Materials, vol. 15, no. 12, 2022, doi:

of
10.3390/ma15124209.

ro
[75] A. Kumar, H. C. Arora, P. Kumar, N. R. Kapoor, and M. L. Nehdi, “Machine learning
based graphical interface for accurate estimation of FRP-concrete bond strength under

-p
diverse exposure conditions,” Developments in the Built Environment, vol. 17, p.
100311, Mar. 2024, doi: 10.1016/J.DIBE.2023.100311.
re
[76] C. Cakiroglu, K. Islam, G. Bekdaş, S. Kim, and Z. W. Geem, “Interpretable Machine
Learning Algorithms to Predict the Axial Capacity of FRP-Reinforced Concrete
lP

Columns,” Materials 2022, Vol. 15, Page 2742, vol. 15, no. 8, p. 2742, Apr. 2022, doi:
10.3390/MA15082742.
[77] K. Turk, C. Kina, H. Tanyildizi, E. Balalan, and M. L. Nehdi, “Machine Learning
a

Prediction of Residual Mechanical Strength of Hybrid-Fiber-Reinforced Self-

rn

consolidating Concrete Exposed to Elevated Temperature,” Fire Technology 2023 59:5,

vol. 59, no. 5, pp. 2877–2923, Jul. 2023, doi: 10.1007/S10694-023-01457-W.
u

[78] O. Alshboul, G. Almasabha, K. F. Al-Shboul, and A. Shehadeh, “A comparative study

Jo

of shear strength prediction models for SFRC deep beams without stirrups using
Machine learning algorithms,” Structures, vol. 55, pp. 97–111, Sep. 2023, doi:
10.1016/J.ISTRUC.2023.06.026.
[79] O. Alshboul, G. Almasabha, A. Shehadeh, and K. Al-Shboul, “A comparative study of
LightGBM, XGBoost, and GEP models in shear strength management of SFRC-
SBWS,” Structures, vol. 61, p. 106009, Mar. 2024, doi:
10.1016/J.ISTRUC.2024.106009.
[80] K. F. Al-Shboul, G. Almasabha, A. Shehadeh, and O. Alshboul, “Exploring the efficacy
of machine learning models for predicting soil radon exhalation rates,” Stochastic
Environmental Research and Risk Assessment, vol. 37, no. 11, pp. 4307–4321, Nov.
2023, doi: 10.1007/S00477-023-02509-X/METRICS.
[81] A. Shehadeh, O. Alshboul, R. E. Al Mamlook, and O. Hamedat, “Machine learning
models for predicting the residual value of heavy construction equipment: An evaluation
of modified decision tree, LightGBM, and XGBoost regression,” Autom Constr, vol.
129, p. 103827, Jan. 2021, doi: 10.1016/j.autcon.2021.103827.
[82] S. Lee, T. P. Vo, H.-T. Thai, J. Lee, and V. Patel, “Strength prediction of concrete-filled
steel tubular columns using Categorical Gradient Boosting algorithm,” Eng Struct, vol.
238, p. 112109, Jan. 2021, doi: 10.1016/j.engstruct.2021.112109.
[83] T. Nguyen-Sy, J. Wakim, Q.-D. To, M.-N. Vu, T.-D. Nguyen, and T.-T. Nguyen,
“Predicting the compressive strength of concrete from its compositions and age using
the extreme gradient boosting method,” Constr Build Mater, vol. 260, p. 119757, Jan.

of
2020, doi: 10.1016/j.conbuildmat.2020.119757.

ro
[84] J. Duan, P. G. Asteris, H. Nguyen, X.-N. Bui, and H. Moayedi, “A novel artificial
intelligence technique to predict compressive strength of recycled aggregate concrete

-p
using ICA-XGBoost model,” Eng Comput, vol. 37, no. 4, pp. 3329–3346, Jan. 2021,
doi: 10.1007/s00366-020-01003-0.
re
[85] B. Zhu, J. Pan, B. Nematollahi, Z. Zhou, Y. Zhang, and J. Sanjayan, “Development of
3D printable engineered cementitious composites with ultra-high tensile ductility for
lP

digital construction,” Mater Des, vol. 181, 2019, doi: 10.1016/j.matdes.2019.108088.

[86] L. Pham, P. Tran, and J. Sanjayan, “Steel fibres reinforced 3D printed concrete: Influence
of fibre sizes on mechanical performance,” Constr Build Mater, vol. 250, p. 118785,
a

Jan. 2020, doi: 10.1016/j.conbuildmat.2020.118785.

rn

[87] O. Alshboul, R. E. Al Mamlook, A. Shehadeh, and T. Munir, “Empirical exploration of

predictive maintenance in concrete manufacturing: Harnessing machine learning for
u

enhanced equipment reliability in construction project management,” Comput Ind Eng,

Jo

vol. 190, p. 110046, Apr. 2024, doi: 10.1016/J.CIE.2024.110046.

[88] F. Althoey et al., “Machine learning based computational approach for crack width
detection of self-healing concrete,” Case Studies in Construction Materials, vol. 17, p.
e01610, Jan. 2022, doi: 10.1016/j.cscm.2022.e01610.
[89] M. H. Graham, “Confronting multicollinearity in ecological multiple regression,”
Ecology, vol. 84, no. 11, 2003, doi: 10.1890/02-3114.
[90] M. D. Manshadi, M. Ghassemi, S. M. Mousavi, A. H. Mosavi, and L. Kovacs,
“Predicting the Parameters of Vortex Bladeless Wind Turbine Using Deep Learning
Method of Long Short-Term Memory,” Energies (Basel), vol. 14, no. 16, p. 4867, Jan.
2021, doi: 10.3390/en14164867.
[91] L. Chen et al., “Development of predictive models for sustainable concrete via genetic
programming-based algorithms,” Journal of Materials Research and Technology, vol.
24, pp. 6391–6410, Jan. 2023, doi: 10.1016/j.jmrt.2023.04.180.
[92] H. M, “3.07 - Robust Calibration,” Comprehensive Chemometrics, pp. 315–343, 2009,
Accessed: Sep. 27, 2024. [Online]. Available:
http://www.sciencedirect.com:5070/referencework/9780444527011/comprehensive-
chemometrics
[93] S. Brown, R. Tauler, and B. Walczak, Comprehensive Chemometrics, vol. 1–4. 2010.
doi: 10.1016/C2009-0-28356-5.
[94] A. Karbassi, B. Mohebi, S. Rezaee, and P. Lestuzzi, “Damage prediction for regular

of
reinforced concrete buildings using the decision tree algorithm,” Comput Struct, vol.

ro
130, pp. 46–56, Jan. 2014, doi: 10.1016/j.compstruc.2013.10.006.
[95] W. Ben Chaabene, M. Flah, and M. L. Nehdi, “Machine learning prediction of

-p
mechanical properties of concrete: Critical review,” Constr Build Mater, vol. 260, p.
119889, Jan. 2020, doi: 10.1016/j.conbuildmat.2020.119889.
re
[96] H. I. Erdal, “Two-level and hybrid ensembles of decision trees for high performance
concrete compressive strength prediction,” Eng Appl Artif Intell, vol. 26, no. 7, pp.
lP

1689–1697, Jan. 2013, doi: 10.1016/j.engappai.2013.03.014.

[97] A. Ahmad, W. Ahmad, F. Aslam, and P. Joyklad, “Compressive strength prediction of
fly ash-based geopolymer concrete via advanced machine learning techniques,” Case
a

Studies in Construction Materials, vol. 16, p. e00840, Jan. 2022, doi:

rn

10.1016/j.cscm.2021.e00840.
[98] C. Ferreira, “Gene Expression Programming in Problem Solving,” in Soft Computing
u

and Industry, Springer London, 2002, pp. 635–653. doi: 10.1007/978-1-4471-0123-

Jo

9_54.
[99] R. Alyousef et al., “Forecasting the strength characteristics of concrete incorporating
waste foundry sand using advance machine algorithms including deep learning,” Case
Studies in Construction Materials, vol. 19, p. e02459, Jan. 2023, doi:
10.1016/j.cscm.2023.e02459.
[100] C. Ferreira, Gene Expression Programming Mathematical Modeling by an Artificial
Intelligence. 2006.
[101] S. W. Wilson, “Classifier Conditions Using Gene Expression Programming,” in
Learning Classifier Systems, E. and B. M. V. and K. T. and L. X. and T. K. Bacardit
Jaume and Bernadó-Mansilla, Ed., Berlin, Heidelberg: Springer Berlin Heidelberg,
2008, pp. 206–217.
[102] H. Adel, S. M. M. Palizban, S. S. Sharifi, M. I. Ghazaan, and A. H. Korayem, “Predicting
mechanical properties of carbon nanotube-reinforced cementitious nanocomposites
using interpretable ensemble learning models,” Constr Build Mater, vol. 354, p. 129209,
Jan. 2022, doi: 10.1016/j.conbuildmat.2022.129209.
[103] G. C. O. Mihai, “A Comparison of Several Linear Genetic Programming Techniques,”
Complex Systems, vol. 14, no. 4, pp. 285–314, 2003.
[104] A. H. Alavi, A. H. Gandomi, M. G. Sahab, and M. Gandomi, “Multi expression
programming: a new approach to formulation of soil classification,” Eng Comput, vol.
26, no. 2, pp. 111–118, Jan. 2010, doi: 10.1007/s00366-009-0140-7.
[105] S. M. Mousavi, A. H. Alavi, A. H. Gandomi, M. A. Esmaeili, and M. Gandomi, “A data

of
mining approach to compressive strength of CFRP-confined concrete cylinders,”

ro
Structural Engineering and Mechanics, vol. 36, no. 6, pp. 759–783, Jan. 2010, doi:
10.12989/sem.2010.36.6.759.

-p
[106] R. Qiu, Y. Wang, D. Wang, W. Qiu, J. Wu, and Y. Tao, “Water temperature forecasting
based on modified artificial neural network methods: Two cases of the Yangtze River,”
re
Science of The Total Environment, vol. 737, p. 139729, Jan. 2020, doi:
10.1016/j.scitotenv.2020.139729.
lP

[107] J. Pyo, S. M. Hong, Y. S. Kwon, M. S. Kim, and K. H. Cho, “Estimation of heavy metals
using deep neural network with visible and infrared spectroscopy of soil,” Science of
The Total Environment, vol. 741, p. 140162, Jan. 2020, doi:
a

10.1016/j.scitotenv.2020.140162.
rn

[108] H. Nguyen, T. Vu, T. P. Vo, and H.-T. Thai, “Efficient machine learning models for
prediction of concrete strengths,” Constr Build Mater, vol. 266, p. 120950, Jan. 2021,
u

doi: 10.1016/j.conbuildmat.2020.120950.
Jo

[109] T. Nguyen, A. Kashani, T. Ngo, and S. Bordas, “Deep neural network with high‐order
neuron for the prediction of foamed concrete strength,” Computer-Aided Civil and
Infrastructure Engineering, vol. 34, no. 4, pp. 316–332, Jan. 2019, doi:
10.1111/mice.12422.
[110] A. H. Gandomi, A. H. Alavi, M. R. Mirzahosseini, and F. M. Nejad, “Nonlinear Genetic-
Based Models for Prediction of Flow Number of Asphalt Mixtures,” Journal of
Materials in Civil Engineering, vol. 23, no. 3, pp. 248–263, Jan. 2011, doi:
10.1061/(ASCE)MT.1943-5533.0000154.
[111] A. H. Gandomi and D. A. Roke, “Assessment of artificial neural network and genetic
programming as predictive tools,” Advances in Engineering Software, vol. 88, 2015, doi:
10.1016/j.advengsoft.2015.05.007.
[112] L. S.-I. L. S. G.G. Erion, “Consistent Individualized Feature Attribution for Tree
Ensembles,” in arXiv preprint arXiv:1802.03888., 2018.
[113] L. S.-I. L. S. M, “A Unified Approach to Interpreting Model Predictions,” in 31st
Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA,
USA., 2017.
[114] H. Alabduljabbar, M. Khan, H. H. Awan, S. M. Eldin, R. Alyousef, and A. M. Mohamed,
“Predicting ultra-high-performance concrete compressive strength using gene
expression programming method,” Case Studies in Construction Materials, vol. 18, p.
e02074, Jan. 2023, doi: 10.1016/j.cscm.2023.e02074.
[115] M. Iqbal, D. Zhang, F. E. Jalal, and M. F. Javed, “Computational AI prediction models

of
for residual tensile strength of GFRP bars aged in the alkaline concrete environment,”

ro
Ocean Engineering, vol. 232, p. 109134, Jan. 2021, doi:
10.1016/j.oceaneng.2021.109134.

-p
[116] M. I. Khan et al., “Effective use of recycled waste PET in cementitious grouts for
developing sustainable semi-flexible pavement surfacing using artificial neural network
re
(ANN),” J Clean Prod, vol. 340, p. 130840, Jan. 2022, doi:
10.1016/j.jclepro.2022.130840.
lP

[117] M. Iqbal, Q. Zhao, D. Zhang, F. E. Jalal, and A. Jamal, “Evaluation of tensile strength
degradation of GFRP rebars in harsh alkaline conditions using non-linear genetic-based
models,” Materials and Structures/Materiaux et Constructions, vol. 54, no. 5, 2021, doi:
a

10.1617/s11527-021-01783-x.
rn

[118] A. Shehadeh, O. Alshboul, K. F. Al-Shboul, and O. Tatari, “An expert system for
highway construction: Multi-objective optimization using enhanced particle swarm for
u

optimal equipment management,” Expert Syst Appl, vol. 249, p. 123621, Sep. 2024, doi:
Jo

10.1016/J.ESWA.2024.123621.
[119] T. Shafighfard, F. Bagherzadeh, R. A. Rizi, and D.-Y. Yoo, “Data-driven compressive
strength prediction of steel fiber reinforced concrete (SFRC) subjected to elevated
temperatures using stacked machine learning algorithms,” Journal of Materials
Research and Technology, vol. 21, pp. 3777–3794, Jan. 2022, doi:
10.1016/j.jmrt.2022.10.153.
[120] M. N. Uddin, J. Ye, B. Deng, L. Li, and K. Yu, “Interpretable machine learning for
predicting the strength of 3D printed fiber-reinforced concrete (3DP-FRC),” Journal of
Building Engineering, vol. 72, p. 106648, Jan. 2023, doi: 10.1016/j.jobe.2023.106648.
[121] H. Izadgoshasb, A. Kandiri, P. Shakor, V. Laghi, and G. Gasparini, “Predicting
Compressive Strength of 3D Printed Mortar in Structural Members Using Machine
 Learning,” Applied Sciences, vol. 11, no. 22, p. 10826, Jan. 2021, doi:
10.3390/app112210826.
[122] A. Ghasemi and M. Z. Naser, “Tailoring 3D printed concrete through explainable
artificial intelligence,” Structures, vol. 56, p. 104850, Jan. 2023, doi:
10.1016/j.istruc.2023.07.040.
[123] M. Alyami et al., “Predictive modeling for compressive strength of 3D printed fiber-
reinforced concrete using machine learning algorithms,” Case Studies in Construction
Materials, vol. 20, p. e02728, Jan. 2024, doi: 10.1016/j.cscm.2023.e02728.
[124] Y. Weng, M. Li, D. Zhang, M. J. Tan, and S. Qian, “Investigation of interlayer adhesion
of 3D printable cementitious material from the aspect of printing process,” Cem Concr

of
Res, vol. 143, 2021, doi: 10.1016/j.cemconres.2021.106386.

ro
[125] Y. Chen, S. He, Y. Zhang, Z. Wan, O. Çopuroğlu, and E. Schlangen, “3D printing of
calcined clay-limestone-based cementitious materials,” Cem Concr Res, vol. 149, 2021,
doi: 10.1016/j.cemconres.2021.106553.
-p
[126] S. Delvasto, A. E. Naaman, and J. L. Throne, “Effect of pressure after casting on high
re
strength fibre reinforced mortar,” International Journal of Cement Composites and
Lightweight Concrete, vol. 8, no. 3, 1986, doi: 10.1016/0262-5075(86)90039-4.
a lP
u rn

Declaration of Competing Interest

Jo

The authors declare that they have no known competing financial interests or personal
relationships that could have appeared to influence the work reported in this paper