Construction and Building Materials 440 (2024) 137393

Contents lists available at ScienceDirect

Construction and Building Materials

journal homepage: www.elsevier.com/locate/conbuildmat

Modeling green recycled aggregate concrete using machine learning and

variance-based sensitivity analysis
Mahmoud Owais a, b, *, 1 , Lamiaa K. Idriss b
a
Civil Engineering Department, Faculty of Engineering, Assiut University, Assiut, Egypt
b
Civil Engineering Department, Faculty of Engineering, Sphinx University, Egypt

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

Keywords: Recycled aggregate concrete (RAC) is commonly used to lessen the environmental effect of concrete building and
Green concrete demolition waste. The compressive strength of the RAC is one of the most critical factors influencing concrete
Global sensitivity analysis quality. The compressive strength is assessed by a compression test, which takes a large number of materials and
Recycled aggregate
is expensive and time-consuming. With the development of novel concrete mixes and applications, academics are
Deep learning
Pretraining machine learning
obliged to seek accurate models for forecasting mechanical strength. A significant source of difficulty in
compressive strength modeling is that there are many mixture components and testing conditions whose vari­
ation significantly influences the predicted values. To this end, this study explores the mixture design of sus­
tainable concrete in order to generate eco-friendly concrete mixes. Tests are conducted on 18 different mixtures
comprising different proportions of waste tires, plastic, cement, and red brick to experiment with new green RAC
mixtures. For the modeling part, the deep residual neural networks (DRNNs) method is first presented to the
problem, aided by a database from the literature for a pretraining task. The proposed DRNNs structure uses
shortcuts (i.e., residual connections) that bypass some layers in the deep network structure to alleviate the
problem of training with high accuracy. The performance of the proposed DRNNs is evaluated using different
goodness of fit measures and compared with well-known machine learning tools. The findings showed that the
suggested model could provide credible predictions about the desired mechanical parameter, saving the required
lab efforts by 40 %. Finally, a variance-based global sensitivity analysis is performed with the Latin hypercube
simulation method to help rank/prioritize each mixture component’s impact on determining the compressive
strength in practice while mitigating the potential misrepresentation of results due to the correlations between
the input parameters. The analysis showed that cement and waste contents are the most significant ones in their
first and total order effects.

1. Introduction and background tied to the structure’s performance and safety throughout its life cycle
[2].
Concrete is one of the materials most often utilized in civil engi­ Concrete is a mixture of cement, water, fine/coarse aggregates, and
neering applications. With about 10 billion tons of concrete produced additional mixtures. This mixture is a complicated system made up of
yearly and expected to reach 18 billion by 2050, it is the most consumed random combinations of various components spread across their nu­
substance in the world [1]. In order to fulfill demand, its worldwide merical domains [3,4]. This heterogeneous trait makes it challenging to
production rate is rising. One of the most crucial factors when devel­ anticipate some mechanical parameters, particularly compressive
oping and assessing concrete is its mechanical qualities. Enhancing the strength [5,6]. Of all the tests conducted on the concrete, the
design of concrete buildings requires a deeper comprehension of con­ compressive strength test is the most crucial one because it provides
crete performance, which is dependent on precise mechanical property information about the material’s characteristics, indicates its ability to
evaluation. Among the many characteristics of concrete, compressive support the weight of the building or structure, and determines how a
strength is thought to be a reliable measure of performance. It is closely material will behave or respond to a compressive load by monitoring

* Corresponding author at: Civil Engineering Department, Faculty of Engineering, Assiut University, Assiut, Egypt.
E-mail address: [email protected] (M. Owais).
1
https://scholar.google.com.eg/citations?user=UNwlx2MAAAAJ&hl=en

https://doi.org/10.1016/j.conbuildmat.2024.137393
Received 30 March 2024; Received in revised form 27 May 2024; Accepted 6 July 2024
Available online 15 July 2024
0950-0618/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

critical parameters like strain, stress, and deformation. The compressive of gypsum, hydrated cement paste, and other contaminants. Given that
strength of concrete after 28 days of aging is the most widely used aggregate comprises the majority of concrete by volume, it is sensible to
measure of its technical qualities and performance, and it is an impor­ look at using leftover concrete as aggregate in newly constructed pro­
tant factor in structural design [7]. The compressive strength after 28 jects [45]. The examined aggregates from the destroyed concrete
days of age is related to other mechanical parameters such as tensile and buildings had comparatively low strengths, and their uses were not very
flexural strength [8]. Furthermore, it is well-known that concrete important.
strength is mostly affected by the water-to-cement-ratio (w/c). However, To this end, this study contributes to the topic in three different di­
even for a given w/c, significant differences in concrete strength may be mensions. First, green recycled aggregate concrete (GRAC) is introduced
seen depending on the properties of the component materials, such as as an environmentally efficient solution, discussing the possibility of
cement type, aggregate type, paste content, mineral and chemical ad­ replacing natural coarse and fine aggregate with new mixes of RA in
mixtures, and so on [9]. As a result, a robust, predictive model capable of structural concrete at different replacement proportions. Second, a new
estimating compressive strength as a function of combination pro­ sophisticated ML tool is used for the RAC modeling problem to learn the
portions would allow for high throughput mixture design while reducing latent relationships among the explanatory variables and their effect on
the empirical, labor-intensive character of "trial batching" methodolo­ compressive strength. The proposed model is pre-trained on literature-
gies that are now used in industry. documented results to reduce the number of experimentation trials for
Physical tests on specimens cured to the specified age are the most new mixture designs. Finally, one of the well-known global sensitivity
straightforward technique to assess concrete’s compressive strength [10, analysis (GSA) tools is used to interpret the interrelated relationships
11]. Such a method of assessing compressive strength takes time and is among the RAC mixture components. GSA will help to rank input factors
influenced by other factors such as specimen production and test oper­ based on their effect/sensitivity in determining the resulting compres­
ators [12]. Furthermore, the test is known to cause specimen deterio­ sive strength. Using ML is a challenging task associated with high
ration. Empirical regression techniques were therefore presented to complexity and uncertainty when it comes to concrete experiments.
predict compressive strength [13–15], but their drawback is the Thus, creating a global predictive model that uses the least amount of
non-linear connection between the compressive strength and its corre­ data becomes necessary. This study attempts to expand on the prior
sponding concrete mixture. This precludes an appropriate prediction knowledge and models gained in concrete mixture prediction utilizing a
accuracy. With improvements in artificial intelligence, the propensity of deep learning architecture. The goal is to minimize the number of testing
applying machine learning (ML) methods to predict the mechanical samples needed for the prediction while maintaining appropriate
characteristics of concrete has gained traction in recent years. ML is a accuracy.
prominent subfield of artificial intelligence concerned with creating and The remainder of the article is as follows. The next section draws the
developing techniques that are eligible for finding complicated patterns state of the art of related topics to highlight the contribution of this
in experimental data without using a preconceived equation [16,17]. In article. Section 3 gives the input data and corresponding factors un­
general, ML aims to create intelligent systems that learn from previous derstudy. Section 4 illustrates the analysis framework with its corre­
knowledge. Data mining (such as online search) and the implementation sponding mathematical and statistical methods. Section 5 presents both
of challenging software systems, such as automated driving, are two modeling and sensitivity results. The final section concludes the
important examples of ML where its models may predict and explain research.
knowledge development based on data [18,19]. ML is applied in struc­
tural engineering in a variety of applications, including compressive 2. The state of the art
strength [20], structural system identification [21], seismic perfor­
mance evaluation [22], vibration control [23], and tensile strength This section is partitioned into three subsections. The first subsection
modeling [24], to mention a few. focuses on reviewing the contributions of previous studies of the con­
From another perspective, the concrete industry has, fortunately, crete mixture to enhance its properties. The second is devoted to ma­
found some sustainable and eco-friendly alternatives to mix Green chine learning tools and their application in predicting the mechanical
Concrete (GC). GC is a contemporary term in concrete mix design in properties of the concrete mix. The third briefly describes the state of the
which the type of concrete has at least one component made of waste, a art of the GSA and its role in modeling problems. Lastly, we will derive
production process that does not harm the environment, high perfor­ the contribution of this study, which has a salient contribution regarding
mance, and life cycle sustainability [25–42]. Most of the studies illus­ each aspect of the stated state of the art.
trated the effect of additional material to coarse recycled aggregate
concrete (RAC) on the properties to improved mechanical properties of 2.1. Advances in concrete mixture
sustainable GC. Otherwise, GC is made with waste material as a partial
or complete replacement for coarse or fine aggregate or cement. The Reference [26] is one of the early studies that examined the perfor­
used materials are blast furnace slag, glass aggregate, concrete aggre­ mance of concrete made using RA, where different durability indexes
gate, recycled demolition waste aggregate, and manufactured sand. were used. The inclusion of RA reduced the concrete mixture perme­
Researchers are using RAC to protect natural resources and produce ability index at a certain curing age. In contrast, for a concrete mix
sustainable structures because of the building industry’s increased need containing a specific quantity of RA, compressive strength mostly de­
for natural aggregates. Nevertheless, waste products are also substituted pends on the original concrete’s strength, which is also influenced by the
for cement in the creation of the GC, including wood ash, recovered recycled concrete’s weight-to-cement ratio. Reference [25] provided
glass, silica fume, fly ash, and blast furnace slag. When making GC, evidence that the kind of concrete matters less than its physical attri­
substitutes for aggregate include recycled concrete materials, waste butes when considering its composition with recycled material. When
plastic, demolition debris, recovered glass aggregate, foundry sand, and the replacement was 20 percent or 50 percent, the carbonation depth in
stone crusher waste. When RAC is used in concrete applications, per­ the concrete with recycled particles dropped. However, the body of
formance flaws such as poor strength and quick collapse under static and research indicates that, under the assumption that all other mix design
dynamic loads are exposed [22]. Research on this topic began with parameters remain constant, the modulus of elasticity decreases as the
fundamental findings regarding the effects of using recycled aggregates amount of recycled aggregate replacement rises. This loss is directly
(RA) on concrete’s compressive strength [43], as well as the practicality related to the lower elastic modulus of RA, which in turn controls the
of doing so economically [44]. Concrete debris produced during con­ final concrete’s compressive strength. It has been determined that the
struction projects makes up a significant amount of the trash volume. modulus of elasticity is not significantly affected by the use of up to 30
They produce pieces where the aggregate is tainted with small amounts percent RAC. However, field and project engineers often employ a

2
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

straightforward formula to get the modulus of elasticity using the characteristics and that the matrix structure became more compact as a
compressive strength of a particular concrete mix while planning and result of the gel products that were continually produced.
constructing buildings. The link between these two qualities shows The possibility of using crushed concrete aggregates in place of
comparable growth even if they appear differently in the current naturally occurring crushed stone aggregates in concrete (CCA) was
structural concrete standards (CEB-FIP, 1990, EN-1992, ACI-318–2008) assessed in [30]. The experimental research takes into account three
[46–48]. variables: the percentage of coarse aggregate replacement, the cement
Numerous studies compared the RAC to the normal aggregate con­ content in concrete, and the w/c. The strength prediction models were
crete (NAC) [27,49–51]. The studies focused on concrete that was cast developed using data from existing literature and the findings of the
using either normal or recycled coarse aggregates and had target current investigation. These models may be used as design data for CCA
compressive cube strengths ranging from 20 to 50 MPa at different w/c concrete in structural members and are useful as basic tools for esti­
ratios with different RAC replacement percentages. According to [49], if mating the strength of concrete at various aggregate replacement levels.
water reducers are used to provide sufficient workability, it is possible to The replacement ratio of particles significantly influenced the strength
boost strength in RAC by decreasing the w/c in NAC. The strength of the characteristics of concrete. Up to 25 % CCA replacement of NAC does
broken concrete that was used to create the RA does not determine the not significantly change the strength characteristics of concrete. Also,
compressive strength of RAC. Furthermore, RAC and NAC had compa­ reference [56] used a mix of experimental and theoretical research to
rable patterns in compressive strength development, with NAC look into the effects of the w/c ratio and RA content on the chlorine
demonstrating a comparatively quicker strength growth up to the age of erosion resistance of RAC. Furthermore, a mathematical equation
seven days. incorporating pore structure development, energy conservation, and
A comparative examination of the experimental findings of fresh and fluid mechanics principles is formulated to depict RAC’s chloride ion
cured concrete characteristics with varying replacement ratios of natu­ transport process. The proposed mixture in [25] was made using fine
ral with recycled coarse aggregate was shown in [27]. Wherever the and coarse RA from destroyed masonry and concrete buildings. The
method used to prepare RA for concrete mixes affects how workable the findings demonstrate that the concrete family with the biggest pore
concrete is; if surface dry, water-saturated RA is utilized, the workability volume did not necessarily match the concrete family with the highest
of NAC and RAC is almost equal. Reference [52] compared the me­ degree of carbonation for compressive strengths of 2900, 4350, and
chanical and durability characteristics of the RAC and the NAC, 5800 psi (20, 30, and 40 MPa). Conversely, concrete produced with RA
including various replacement percentages for RACs (e.g., 0 %, 10 %, (20 %, 50 %, and 100 % replacement) from old masonry or concrete may
15 %, 20 %, 25 %, 50 %, and 75 %). In the comparison of the test results attain the same compressive strength at 28 days, in the range of
reported cube compressive strength, the RAC specimens at the observed 20–40 MPa, as well as the same fresh workability as NAC.
ages were marginally weaker than the NAC specimens. While the per­ Reference [31] used glass debris that had been milled as a supple­
centage growth of compressive strength for the RAC and NAC specimens mentary cementitious material to create RAC that had better strength
had comparable patterns, the NAC specimens exhibited a comparatively and durability. Waste glass, ground to micro-scale particle size, is ex­
quicker strength rise over the first seven days of curing. pected to conduct pozzolanic reactions with cement hydrates, resulting
Reference [53] showed potential pozzolan slurry – which comprises in secondary calcium silicate hydrate. The new notion of employing
fly ash, nano-SiO2, and silica fume slurries – and CO2 treatments were milled waste glass as a partial substitute for cement was tested experi­
looked at as RAC enhancing techniques. According to test findings, mentally to address the disadvantages of RA and the resultant concrete.
pozzolan slurry treatment may cause fluidity to diminish, while CO2 These activities improve the structure of the hydrated cement paste as
treatment was more successful in lowering water absorption and well as the interfacial transition zones in RAC. It was discovered that
increasing fluidity. They concluded that the use of treated RAC clearly milled waste glass inhibited alkali-silica reactions. Moreover, using
increased the durability performance and compressive strength. pulverized waste glass instead of some of the cement in RAC leads to
Regardless of the curing age, the compressive strength of the observed improved durability properties like sorption, chloride permeability, and
carbonation depth and chloride migration coefficient was discovered to freeze-thaw resistance. The development of a denser and less permeable
have a nearly linear relationship. Reference [28] showed that treated microstructure, which is anticipated to be the outcome of the filling
RAC with 0.2 percent nano-silica slurry improves the mechanical action of sub-micron-sized glass particles, results in a significant in­
properties of the RAC. Reference [54] determined that the concrete with crease in strength at a later age. Nevertheless, reference [38] used dis­
RA from stone factories has a final compressive strength that is about carded tire rubber in place of aggregates with two different kinds of
9 % higher than the concrete with RA from building trash. Nevertheless, cement, namely sulfate-resistant cement (SRC) and regular Portland
it was reported a 30 % replacement rate for aggregate materials is ideal. cement (RPC). It was discovered that the SRC kind could raise the
Meanwhile, reference [29] investigated the potential of metakaolin as a compressive strength more than RPC by a percentage of 25 %. Alter­
mineral addition to improve the performance of concrete mixes with natively, [57] experimented with three different kinds of cement (CEM
different ratios of RA. Fourteen different concrete mixtures have indi­ III/B 42.5-N-LH/SR, CEM II A/L 42.5/R, and CEM II A/S 42.5 N/SRC)
cated that 15 % of metakaolin could increase compressive strength, for the RAC. The experimentation determined properties of the con­
regardless of the kind of aggregate. This result was affirmed in [32], cretes’ drying shrinkage, chloride permeability, and accelerated
where regular Portland cement was substituted with nano metakaolin in carbonation coefficient in comparison to values estimated by specific
the RAC production. The findings demonstrated that the 28-day standards of exposure to environments. The RAC made with CEM III/B
compressive strength progressively enhanced with the rise in nano has very low chloride ion penetrability, ranging from 500 to 740 Cou­
metakaolin content while the replacement rate of RAC remained con­ lombs. Reference [58] employed carbonation-created treatment to in­
stant, but that the increasing trend eventually slowed down. In addition, crease the qualities of RAC. The impact of RA was examined on the
reference [55] investigated the impact of curing age and w/c on the quality of chloride permeability, the chloride permeability, and steel
microstructure, mechanical characteristics, and damage mechanism of corrosion that results from using fine and coarse carbonated RA in
modified RAC containing metakaolin. The outcome showed that there concrete. Carbonation treatment is found to enhance RA’s physical
was a negative correlation between w/c and the mechanical qualities of performance, raising the prepared RAC’s compressive strength.
the specimens but a positive correlation with curing age. Furthermore, Compared to concrete made with non-carbonated, the steel corrosion
the addition of metakaolin resulted in an average rise of 17.46 percent risk of reinforced RAC is decreased, and a linear relationship is seen
and 8.42 percent at six ages in the specimens’ peak stress and peak strain between the steel corrosion risk and chloride permeability.
with w/c of 0.65. By using microscopic testing, it was discovered that the Reference [37] looked into the characteristics and microstructure of
hydration process was connected to the specimen’s mechanical carbonated coarse RAC. The concrete containing 50 % carbonated

3
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

coarse RAC had a 16.7 % greater compressive strength than uncarbo­ properties while significantly enhancing fracture energy values by up to
nated coarse RAC. More considerable autogenous shrinkage and approximately 16.6 % [67]. However, higher steel slag content may
compressive strength were obtained with a higher carbonated coarse adversely affect relevant properties due to its lower cementitious ac­
RAC replacement ratio. Using a cement paste dissociation agent (CPDA) tivity compared to slag. To mitigate these drawbacks, limiting the steel
and various mixing techniques, reference [59] investigated the RAC slag content to less than 25 % is recommended, ensuring optimal me­
coating treatment that was recently added to RAC concrete mixes. A chanical and fracture properties of GFRAC.
preliminary test program was conducted before determining the These recent findings highlight the importance of ongoing research
appropriate dose of the CPDA solution based on the mechanical strength and development efforts in optimizing the performance of green con­
characteristics of the RAC test findings. Following that, a number of crete, particularly in the context of recycled aggregates and supple­
experimental tests were conducted to examine the effects of RAC coating mentary cementitious materials. By addressing challenges such as
treatment, various mixing techniques, including the equivalent mortar mixture design methodology and material selection, the industry can
volume method and conventional technique, and various RAC replace­ further advance towards sustainable construction practices and reduce
ment ratios on the drying shrinkage and rapid chloride penetration test its environmental footprint.
values of RAC concrete. Adding 2.5 percent of CPDA to the RAC concrete
mixes enhanced the mechanical strength characteristics, including 2.2. Modelling attempts of concrete mixture
elastic modulus, flexural strength, and compressive strength. Further­
more, reference [36] examined surface-coat aggregate with pozzolanic Conventional artificial neural network (ANN) is considered one of
materials and further improved the interfacial transition zone’s micro­ the most used ML techniques in modeling material properties. It is a
structure (ITZ) and RAC characteristics. The impact of surface-coating concept derived from the biological nervous system that processes data
RA in the RAC with different admixtures on ITZ microstructure was like the brain does [68]. The unique structure of the information pro­
also investigated by the preparation of the new-old-new concrete cessing system is a key aspect of this approach, which comprises
sandwich specimen. The mechanical strength characteristics of RAC are numerous highly linked processing parts that collaborate to deal with
improved when fine-grounded slag is used as an admixture instead of fly modeling problems. Each neuron’s output is multiplied by weight co­
ash. However, reference [60] indicated that types of cement-containing efficients and applied to the input power via a nonlinear activation
additives (silica fume, slag, fly ash, etc.) could offset the impact of function [69,70]. Reference [71] employed ANN to forecast the
adding recycled materials. Also, plasticizers were used on reclaimed compressive strength of typical concrete after 28 days. The model
concrete to enhance its mechanical qualities or workability. demonstrated the capacity to determine the compressive strength of
Alternatively, reference [61] used multiwall carbon nanotubes to various concrete mixes. Several studies used various ANN approaches to
enhance these RC characteristics (MWCNT). This work evaluates RAC, test the strength of concrete incorporating RA. In this context, Reference
which has been amended with varying concentrations of MWCNT for its [72] examined the compressive strength of concrete incorporating
fresh and hardened characteristics. The MWCNT greatly improves the construction and demolition debris. It was found that ANN can forecast
RAC specimens’ impact resistance, with enhancements at the first crack compressive strength in both testing and training at 3, 7, 28, and 91 days
and failure phases. Among the various nano levels, the dose of 0.1 of curing. Reference [73] used the ANN method to assess the compres­
percent MWCNT offers the largest percentage gain in impact resistance. sive strength of concrete using RA (coarse aggregate only) with 14
Reference [62] reported improvements in mechanical characteristics various inputs and found that this method may be a helpful tool for
and microstructural alterations of self-compacting RAC, which is made forecasting RAC compressive strength. In addition, reference [74] used
using debris from construction and demolition. Experimental research the ANN structure with 6 inputs and 18 hidden nodes to assess the
shows that silica fume may substitute fly ash by 7 % while still yielding compressive strength of concrete, including construction debris (coar­
the necessary new and enhanced mechanical characteristics. se-grained alternative), as well as the influence of each input factor. The
Reference [63] investigated the use of micro-carbon fiber (CF) in fly findings showed that water absorption has the most considerable in­
ash geopolymer incorporating fine RAC to enhance the material me­ fluence on the strength when compared to the other inputs. Reference
chanical characteristics. The geopolymer mortar mix containing 100 % [75] estimated high-strength concrete’s elastic model and tensile
RAC had higher compressive and breaking tensile strengths when 0.2 strength incorporating RA. It was found that the ANN may predict the
percent CF was added. With a considerable increase in both surface mechanical qualities of RAC without taking into account the process of
abrasion resistance and flexural strength, the best results for these pa­ concrete failure or defining the precise functional link between the in­
rameters were achieved when 50 % RAC was used. On the same track, dependent and dependent variables. Reference [76] utilized ANN to
reference [64] studied the compressive characteristics of fiber-recycled forecast the compressive strength of concrete containing RAC and
aggregate concrete. A uniaxial compressive test was performed on a discovered that this technique is viable for predicting compressive
total of 75 specimens with varying replacement rates of coarse RA and strength. Reference [77] employed ANN to forecast the compressive
fiber type. The inclusion of polyvinyl alcohol had the best increase in the strength of the SCC with bottom ash. Reference [78] used ANN to
RAC’s compressive strength and energy dissipation capacity. The forecast the strength of concrete, incorporating 168 datasets from the
simultaneous impacts of polypropylene fibers and RAC on the work­ literature taking into account different types and sources of RA. Refer­
ability of self-consolidating concrete (SCC) were shown in [35]. Refer­ ences [76,79] showed that the ANN model outperformed decision tree
ence [65] used fiber polymer to enhance the durability of RAC in models and linear regression analysis in estimating the RAC compressive
addition to its mechanical characteristics. Two types of RAC mixtures strength. Reference [80] tested RAC incorporating fly ash and blast
were prepared: RAC reinforced with polypropylene fibers (PPs) and RAC furnace slag using a genetic algorithm to fine-tune the ANN. The simu­
reinforced with basalt fibers (BFs). Although adding BFs showed a lation results suggest that the ANN model is more flexible, capable, and
slightly better impact than PPs, 0.2 percent of PPs and BFs were enough accurate in forecasting the compressive strength of concrete. Further­
to achieve the greatest flexural strength. more, references [81,82] used the ANN to estimate the compressive
Despite its potential benefits, the widespread adoption of GRAC has strength of eco-friendly geopolymer concrete at 7, 28, and 90 days. It
been impeded by the lack of an effective mix design methodology [66]. was finally stated that the model produced good results. In another
Recent studies have aimed to address this challenge, focusing on opti­ research direction, reference [68] predicted the shear strength of
mizing the composition of GRAC to enhance its mechanical properties RAC-reinforced beams using the ANN approach and obtained satisfac­
and durability. Incorporating an appropriate amount of steel slag, tory results. The use of ANNs in reinforced concrete modeling has also
approximately 25 %, in geopolymer fully recycled aggregate concrete been investigated in recent years. These include reference [83] work on
(GFRAC) has been shown to have minimal impact on mechanical the reinforced concrete beams’ shear capacity. Reference [17] used four

4
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

criteria (w/c ratio, concrete compressive strength, specimen age, and [1].
fiber reinforcing index) to estimate the tensile strength of steel In recent years, support vector machine (SVM) modeling techniques
fiber-reinforced concrete. They employed the ANN approach and re­ have been extensively used in a variety of domains to investigate the
ported that it outperformed other ML tools. For ANN usage in reinforced mechanical characteristics of concrete. Researchers have employed
concrete, we recommend these references for further reading: [84–88]. complex, hybrid, and independent models to accomplish their objec­
To enhance ANN prediction performance, reference [89] investi­ tives. For example, when using SVM for regression analysis, the model is
gated the compressive strength of concrete containing smooth soot using often termed support vector regression (SVR). Reference [108] studied
a mix of ANN and multi-objective grey wolves methods (GWM), the compressive strength of RAC incorporating recycled rubber using the
concluding that the combined model was both predictable and accurate. SVR. The research included three distinct kernels (Gaussian, poly­
The same conclusion was obtained when reference [90] employed the nomial, and linear). It was determined that using the L1QP optimization
GWM method in the ANN training phase to estimate the compressive technique, a Gaussian kernel function, and standardized data gave more
strength of regular and high-performance concrete (HPC). In addition, accurate predictions than other SVR models. Reference [95] used the
reference [91] predicted HPC compressive and tensile strength using the SVR model to accurately and efficiently predict the compressive strength
ANN and a modified firefly method (MFA). The findings revealed that of the RAC using RA replacement percentages as inputs. In addition,
the MFA-ANN hybrid system has a minor inaccuracy in predicting reference [109] brought an improved form of SVR, known as least
compressive strength while significantly reducing computation time. squares SVR (LSSVR). The LSSVR was used to estimate the compressive
Meanwhile, reference [92] compared the fuzzy logic (FL) and ANN to strength of SCC. Reference [110] also employed the LSSVR to forecast
forecast the compressive strength of RAC containing silica fume. It was the resilient modulus of concrete, including two kinds of RA (mixed and
stated that both strategies were successful and that the ANN method separate), and determined that LSSVR had the optimum capacity.
outperformed the FL method in terms of coefficient of determination. On Reference [112] upgraded the LSSVR with metaheuristic optimization
the contrary, when reference [93] employed the ANN and the FL to and used it to estimate the compressive strength of the HPC. Reference
forecast the compressive strength of geopolymer concrete with fly ash [111] reported that kernel-based techniques produced good results.
and slag, it was found that the FL model performed better than the ANN Recent research has also looked at the prediction of compressive
in terms of a number of accuracy measures including the coefficient of strength in geopolymer concrete [112–114]. References [104,115] used
determination one. Also, reference [94] employed the ANN and the FL SVR to examine the compressive strength of the HPC, using inputs
techniques to forecast concrete’s compressive and tensile strength using including the primary concrete components and sample age. The accu­
RA after 3, 7, 14, 28, and 90 days. Ref. [94] concluded that the outcomes racy score in both studies showed that the SVR has good prediction
represented by FL and ANN models are similar to the experimental re­ accuracy. Reference [116] applied the SVR-based prediction technique
cords. Furthermore, reference [95] employed convolutional neural to a more challenging issue: the unconfined compressive strength of
networks (CNN) to calculate the compressive strength of concrete built cockle shell "cement" sand mixes.
using RAC. They discovered that the model utilizing CNN outperformed Evolutionary computational (EC) methods [117–120] are one of the
those using traditional neural networks. most cost-effective and straightforward problem-solving methodologies
Decision trees (DT) have been one of the most extensively used data in the area of artificial intelligence (in terms of computing load and
mining approaches during the past two decades. These approaches may algorithm implementation time). EC is defined as a set of general opti­
be used to identify and extract information from databases and build mization algorithms inspired by biological evolution processes [117,
prediction models [96–98]. It may convey its predictions as a set of 121–123]. EC techniques utilize evolutionary processes such as muta­
rules. It typically comprises inner, terminal, and root nodes. The root tion and others; in other words, they control a vast number of poorly
node includes all the data and is separated into two sub nodes on the constructed responses in an iterative process until the system can solve
right and left, each representing an independent variable. Branches the issue with the required accuracy [124–127]. The revised ML ap­
indicate a distinct set of variables. DT-based algorithms have been proaches have limitations when compared to evolutionary algorithms.
proposed as a quick and efficient way to forecast the mechanical char­ Reference [128] investigated the RAC compressive strength with silica
acteristics of concrete. Reference [99] used the M5 tree model to fume by gene expression programming (GEP), where the highest level of
calculate the elastic modulus, tensile strength, and compressive strength agreement between experimental and GEP model findings was reported.
of concrete incorporating RAC and compared it to other approaches. The Reference [129] investigated GC’s elastic modulus, tensile strength, and
M5 model successfully predicts the mechanical characteristics of the compressive strength, including waste foundry sand, using the GEP. The
RAC after correcting for crucial factors. Several studies have also looked findings from different goodness of fit measures in all three learning,
at concrete performance utilizing the M5P algorithm (i.e., modified M5 validation, and testing sets validate the model’s correctness and capa­
model). Reference [100] employed the M5P method to analyze the bilities. Reference [130] found that the GEP model worked well in
elastic modulus of concrete using RA. It was stated that the tree model estimating the shear strength of FRP-reinforced concrete beams. Refer­
built by M5P is an adequate predictor. Reference [101] tested concrete ences [131,132] revealed that the GEP model accurately estimated the
incorporating leftover foundry sand and determined that the M5P al­ shear strength of reinforced concrete beams. In a different work, refer­
gorithm produced appropriate models. References [79,102], and [103] ence [133] used the linear genetic programming (LGP) model to explore
employed concrete ingredients and curing age as input and eventually the compressive strength of reinforced plastic carbon fiber concrete. The
reported satisfactory results using M5P to forecast compressive strength. findings revealed that LGP could forecast compressive strength with a
In this context, reference [104] investigated compressive strength using reasonable accuracy and outperforms numerous models given in previ­
another decision tree-based method (MART) and showed that it suc­ ous research. Alternatively, reference [134] used the GEP to anticipate
cessfully predicted compressive strength with varying HPC ages. In eco-friendly concrete’s compressive and electrical properties. The find­
another approach, reference [105] evaluated the reinforced concrete ings show that, based on statistical characteristics, the GEP can accu­
beam-column joints’ shear strength using a Random Forest (RF) method. rately estimate electrical resistance and compressive strength.
Reference [106] also employed the RF technique to estimate the Ref. [135] projected HPC compressive strength by the GEP. In the di­
compressive strength of the HPC, finding that representing the input rection of reinforced concrete, Ref. [136] calculated the short rectan­
parameters in absolute mass improved model accuracy. In [107], the gular reinforced concrete columns’ shear strength using two nonlinear
compressive strength of lightweight SCC was investigated using a regression and GEP models. Validation of the models revealed that the
combination of beetle antennae search (BAS) and RF algorithm, showing average error for the GEP model was 15 %, whereas the average error for
a strong correlation coefficient for the BAS-RF model. Among the the nonlinear regression was 50 %. Ref. [137] reported the high accu­
different ensemble learning methods, the most used is adaptive boosting racy of the genetic programming method in predicting rice husk ash

5
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

concrete compressive strength. Ref. [138] projected the elastic modulus Variance-Based Sensitivity Analysis [150], and Density Sensitivity
of SCC using two models: artificial bee colony programming (ABCP) and Analysis [151], may be used. These methods depend on indicators that
biogeographical-based programming (BBP). It was found that the BBP range from the correlation between the components to the statistical
model was marginally more similar to the experimental data than the characteristics of the distribution of the underlying variables (e.g.,
ABCP model. Furthermore, the sensitivity of BBP parameters demon­ variance). Except for partial derivatives, the majority of these models’
strates that the BBP model’s prediction increases when maximum tree indices are difficult (if not impossible) to estimate analytically [152,
depth, colony size, and habitat size increase. 153]. Sampling algorithms such as the Latin Hypercube (LHS) or Monte
Practically, most prior research employed single ML algorithms to Carlo (MC) will be used to assess the computation of these indices [154].
predict concrete compressive strength using RAC. Other studies Currently, GSA has the most diverse applications, including model
compared two or more ML tools in their investigation. The flexural, verification and calibration, uncertainty reduction, robust decision-
tensile, and compressive strength of geopolymer SCC was examined in making, and system controller analysis [155–158]. This is owing to
[139] using ANN and GEP models. The findings reveal that both models the fact that it overcomes the limitations of the LSA method. To the best
can predict the mechanical characteristics of concrete with a confidence of our knowledge, GSA is absent from the RAC mixture analysis.
level of around 97 %. Ref. [140] investigated the accuracy of several Therefore, applying the VBSA in this study is one of the contributions
data mining algorithms in predicting the compressive strength of GRAC that will be considered.
(i.e., green recycled aggregate concrete). Ref. [141] evaluated the ten­
sile strength of concrete using RA using three models: GP, ABCP, and 2.4. This study’s contribution
BBP. The models were described as trustworthy techniques for fore­
casting. Ref. [142] projected the lightweight foamed concrete The present work investigates how to create a more sustainable GC
compressive strength using the extreme learning machine (ELM) model throughout its life cycle and outperforms conventional concrete. GRAC
along with other models (e.g., SVR and M5). In this aspect, ELM out­ produces less CO2 emissions and uses less energy to manufacture [159].
performs the other authorized techniques in terms of accuracy and The use of locally sourced and recycled resources always has issues with
reliability. Refs. [104,143] studied several machine learning algorithms workability, shrinkage and creep amounts, and heat of hydration [160,
for forecasting concrete compressive strength, including SVR and ANN. 161]. The present research aims to improve these attributes since the
Ref. [144] employed the Gaussian process and gradient boosting (GB) to acceptability of using RA instead of natural aggregate is mostly deter­
estimate the RAC compressive strength, and it was found that GB had the mined by the properties of the RA [162]. We aim to evaluate and explore
greatest prediction accuracy. the RA used in GC applications for waste tires, waste plastic, and red
Almost all of the above-mentioned published research employed brick at w/c ratios to raise the compressive strengths of the GC.
accuracy measures to assess the model’s predictive accuracy. However, Crushed-weight recycled bricks (CWRB), cutter waste plastic bottles
accuracy metrics alone may not be sufficient to measure the prediction (CWPB), waste rubber tires (CWRT), and crushed waste cement blocks
accuracy of the models; additional indication, such as feature signifi­ (CWCB) are experimented with in the proposed GRAC mixtures to
cance, is also critical to accurately evaluate the model’s prediction ac­ reduce production costs while increasing concrete durability.
curacy [145,146]. This will be addressed in the subsequent section. The deep residual neural networks (DRNNs) method is selected in
this study as the ML tool for the modeling part. Although the DRNNs has
2.3. Global sensitivity analysis proved its high suitability and efficiency in material properties predic­
tion in recent studies [163–165], it has never been used in concrete
Sensitivity Analysis (SA) is typically used as a post-processing step mixture modeling. Its capability to transfer learning from other lab ex­
after measuring or predicting a variable. Because conducting SA in the periments to the current lab is proved in Ref. [164]. Data from the
laboratory is challenging due to the growing demand for samples, the literature is collected for this purpose. The trained DRNNs would be
bulk of SA for material characteristics is done during the modeling considered as the virtual lab to conduct the GSA stage, where numerous
phase. As a result, the results derived from any SA technique rely heavily experiments (i.e., runs) are required. For the post-processing step, the
on the model’s accuracy and capacity to replicate the input factor’s Variance-Based Sensitivity Analysis (VBSA) is selected as the GSA tool.
impact on the output variable. In general, SA examines how a numerical The VBSA has also proved its distinct abilities in many prediction
model’s changes in variables (input factors) affect its response variable problems [166,167] without being called in the concrete mixture design
(output factor) variation. The modeling domain controls the SA’s analysis.
complexity and aims within this broad framework. Throughout the SA To this end, this study contains salient contributions to the literature
application, the analyst must make multiple choices, starting with the that are summarized in the following points:
choice of the output(s) modeling approach, followed by the choice of the
input factor domain limits, how to move the input factors in the domain • Investigating new mixture design for the GC field.
space, and the methodology for the SA [147]. • Applying the DRNNs as a new modeling ML tool.
SA is classified as local (LSA) or global (GSA) according to the • Transferring knowledge from the literature to the current practice.
domain of input factor change. LSA aims to measure the output factor’s • Ranking concrete mixture components regarding their impact on
sensitivity by moving the input factor away from a reference or ideal compressive strength using the VBSA.
value while keeping the other components at nominal levels (e.g., the
mean). In contrast, GSA analyzes the same sensitivity while taking into 3. Input data
consideration the variation throughout the input factor’s whole space
[148]. Furthermore, GSA can handle scenarios in which the other input 3.1. Laboratory experimentation
components are either constant or changing regarding the factor under
examination. In other words, LSA can only study the effect of "one-­ All materials used for the GRAC in this study experimentation are
variable-at-a-time" (OAT), while GSA may analyze the OAT or all com­ produced in Egypt, and the experiments are done in the concrete
ponents simultaneously "all-variables-at-a-time" (AAT). GSA may thus structure laboratory at Sphinx University, as shown in Fig. 1. The
investigate the interactions between the input variables, but LSA cannot. selected cement is SRC [CEM IV/A (P) 52,5 N – SR], which has high
The last decision that must be taken is which methodological durability and compressive strength under aggressive conditions and
methodology will be used to answer the targeted SA questions. Meth­ protects structures from sulfate attack. Furthermore, its very low heat of
odologies ranging from traditional partial derivatives to more contem­ hydration helps to avoid shrinkage cracks. The cement is tested ac­
porary approaches, such as the Elementary Effects test [149], cording to Egyptian standards for cement [E.S.S (4756/1/2013)] and [B.

6
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 1. Green recycled aggregate concrete (GRAC) materials.

S.EN (197–1/2011)] standards [168]. Silica Fume (SF) is used as an In the experimentation stage, three types of CWPB, CWRT, and
additive material, where SF was added as a cementation material in the CWCB composites were prepared in different percentages of 20 %, 40 %,
concrete mix. Also, Superplasticizer Sikament-163 M (S-163 M) was and 60 % content of coarse aggregates with 100 % CWRB content as fine
added to speed hardening early and also increase ultimate strengths. aggregate for all the GRAC mixes. The w/c ratios used in the mixtures
As specified by EN 197–1, the SRC is designed for reinforced or were in different proportions of 0.55, 0.75, and 0.95. The control sam­
prestressed concrete applications requiring high initial and final ples with RAC were obtained by crushing the pre-tested cubes.
strength. It is ideal for chemically aggressive environments, such as Accordingly, 18 mixes were constructed to study the compressive
those exposed to sulfate attack or seawater, enhancing the durability strength of the proposed GRAC components. The lab experimental mixes
and longevity of concrete structures. This cement resists chloride ion are summarized in Table 3. Cement, RFA, and RAC were mixed in the
penetration, preventing corrosion of reinforcing steel and ensuring dry state by hand, and then the required quantity of water was added
structural durability. It is engineered to withstand high sulfate concen­ and mixed thoroughly, as shown in Fig. 2. After casting the concrete
trations, making it suitable for foundations, sewage treatment plants, cubes, these specimens were kept in the molds for 24 hours. After that,
and marine structures. Additionally, its low heat of hydration reduces these specimens were cured in curing basins for 7, 14, and 28 days. In
thermal cracking in large pours, contributing to long-term structural addition, the GRAC mix specimens were prepared ↱based on recipes
stability. The cement’s slow hydration process allows for continued developed in accordance with the EN 12390 ↱standard and underwent
strength gain and a denser microstructure over time, ensuring the tests that evaluated the following properties: ↱workability and
durability of concrete in severe environmental conditions. This combi­ compressive strength.
nation of properties makes the SRC essential for constructing durable,
sustainable infrastructure projects [169,170]. 3.2. Literature database
SF and Sikament were produced by Sika Company in Egypt, and they
conform with A.S.T.M. C 494–92 Type F. All properties of each cement In this subsection, the database collected from the literature is
and SF are demonstrated in Table 1 and Table 2. The physical and me­ summarized in Table 4. Thirty authenticated studies from the reviewed
chanical properties of S-163 M by the manufacturer datasheet are as literature are used to collect 311 concrete compressive strength data
follows: density at 200 C = 1.20 kg/liter, pH value = 4.3–4.7, and the points at 28 days of curing. The studies are selected based on their
material state is light brown liquid. Fine aggregate is used of crushed credibility (i.e., published in indexed journals) and the availability of the
waste red bricks (CWRB) to replace sand. The fineness modulus of the selected input data variables. It is worth noting that the input variables
CWRB is 2.60. The micro powder size of the CWRB is smaller than are also carefully chosen globally to manage, including possible RAC
0.15 mm. Coarse aggregates are selected among the selected different and GRAC data points. The input variables are categorized into four
percentages of cutter waste rubber tires (CWRT), cutter waste plastic main groups: aggregate characteristics, cement characteristics, additives
bottles (CWPB), and crushed waste cement blocks (CWCB). The average type/quantity, and testing conditions. This study considers one target
coarse aggregates’ properties are of size 5–20 mm, specific gravity = output factor, where the proposed framework application to any other
1.10, and density = 0.53 t/m3, whereas the water absorption for the concrete property is a straightforward task. Numerically, the factors
CWRT is 7.80 %. could be classified as continuous (i.e., scalar) and discrete (i.e., cate­
gorical). The statistics of scalar variables are presented in Table 4, where
Table 1 the categories of the discrete variables are presented in the table notes.
The properties of SRC. The detailed discussion of the types of discrete variable categories is
omitted for the sake of a concise presentation since it is beyond the scope
Specific % of water Specific surface Bulk Soundness
Gravity gm/ consistency area (Bline) density (Le of this article. It is easy for researchers and practitioners to track the
cm3 cm2/g (gm/ Chatelier) mentioned types in the reviewed literature.
cm3) mm

3.14 28 % 4066 1.14 0.15 4. The analysis framework

Initial setting Final setting Compressive strength (MPA)
time (min) time (min) 7 Days 14 Days 28 Days The developed framework consists of four steps: data preparation,
145 195 37 50 52.5
modeling, testing, and post-processing analysis. The initial phase

7
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Table 2
Physical and mechanical properties of silica fume by the manufacturer datasheet.
Specific Mean Specific Bulk Silicon Aluminum Calcium Magnesium Potassium Sodium Sulfur Water color
Gravity grain surface area density dioxide oxide AL2O3 oxide CaO Oxide MgO oxide K2O oxide dioxide H2O
size (m2/kg) (gm/cm3) SiO2 Na2O SO2

2.15 7.00 178,000 3.45 96 1.10 1.20 0.18 1.20 0.45 0.45 0.85 Light
Gray

Table 3
Mix proportions of the GRAC.
Mix NO. Mix. Code Cement Natural sand (kg/ CWRB (kg/ RAC CWPB (kg/ CWRT (kg/ (CWCB) (kg/ Water silica Plasticizer
(GRAC) (kg/ m3) m3) (kg/ m3) m3) m3) (kg/ fume (kg/m3)
m3) m3) m3) (kg/m3)

Control RAC SRC396 804 ———— 423 0 423 0 217 39.6 12

GRAC1 20 P-20 T-60B- SRC396 0 804 0 168 168 510 217 39.6 12
0.55
GRAC2 20 P-40 T-40B- SRC396 0 804 0 168 339 339 217 39.6 12
0.55
GRAC3 20 P-60 T-20B- SRC396 0 804 0 168 510 168 217 39.6 12
0.55
GRAC4 20 P-20 T-60B- SRC396 0 804 0 168 168 510 297 39.6 12
0.75
GRAC5 20 P-40 T-40B- SRC396 0 804 0 168 339 339 297 39.6 12
0.75
GRAC6 20 P-60 T-20B- SRC396 0 804 0 168 510 168 297 39.6 12
0.75
GRAC7 20 P-20 T-60B- SRC396 0 804 0 168 168 510 376 39.6 12
0.95
GRAC8 20 P-40 T-40B- SRC396 0 804 0 168 339 339 376 39.6 12
0.95
GRAC9 20 P-60 T-20B- SRC396 0 804 0 168 510 168 376 39.6 12
0.95
GRAC10 40 P-20 T-40B- SRC396 0 804 0 339 168 339 217 39.6 12
0.55
GRAC11 40 P-40 T-20B- SRC396 0 804 0 339 339 168 217 39.6 12
0.55
GRAC12 40 P-30 T-30B- SRC396 0 804 0 339 254 254 217 39.6 12
0.55
GRAC13 40 P-20 T-40B- SRC396 0 804 0 339 168 339 297 39.6 12
0.75
GRAC14 40 P-30 T-30B- SRC396 0 804 0 339 339 168 297 39.6 12
0.75
GRAC15 40 P-30 T-30B- SRC396 0 804 0 339 254 254 297 39.6 12
0.75
GRAC16 40 P-20 T-40B- SRC396 0 804 0 339 168 339 376 39.6 12
0.95
GRAC17 40 P-40 T-20B- SRC396 0 804 0 339 339 168 376 39.6 12
0.95
GRAC18 40 P-30 T-30B- SRC396 0 804 0 339 254 254 376 39.6 12
0.95

RAC: recycled aggregate concrete, GRAC: green recycled aggregate concrete, CWPB: cutter waste plastic bottle, CWRT: cutter waste rubber tires,
CWCB: crushed waste cement blocks, CWRB: crushed waste red brick.

involves preparing the data and identifying the relevant components Instead of symbolic techniques, it makes use of statistical and proba­
and their domain areas, as shown in Section 3. In the second stage, as the bility theory tools and models [171,172]. In reality, ML techniques
experimental data is processed as literature data categories, the com­ enable computers to gain the knowledge required to accomplish a
bined data is fed into the DRNNs model, leaving testing data from both certain job by examining a large number of data samples [173,174].
experimental and literature. Left experimental data is used to evaluate Before utilizing the method, a phase known as feature extraction is
the model capability of transfer learning through different labs. The required, which involves extracting the characteristics that provide the
third stage is concerned with testing the model with varying metrics of most specific information. The sample data is used to teach the system to
goodness of fit to build confidence in the GSA. In the last stage, the convey characteristics and differentiate patterns [175–177]. Deep
model is ready to forecast the compressive strength values for a wide Learning (DL) approaches were proposed to address the issues with
variety of synthetically produced input data using the preset factors hand-crafted features in sophisticated ML applications [176]. In-depth
domain (i.e., input data description). The VBSA (i.e., the post-processing learning is motivated by discoveries in neuroscience and is compatible
tool) generates sensitivity results/indices. These phases may be followed with the interpretation of information processing and communication
using the general workflow presented in this research, as shown in patterns in the nervous system. DL layers consist of an artificial neural
Fig. 3. network’s hidden layers and a collection of sophisticated formulae.
Furthermore, based on the training resources, ML is classified as
4.1. Machine learning technology supervised, unsupervised, semi-supervised, or reinforced [178,179].
Supervised and unsupervised learning are the most popular kinds of
ML emerged as a subfield in artificial intelligence in the 1990s. machine learning in a variety of fields, including engineering [180].

8
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 2. Tools and methods for preparing the GRAC samples.

Table 4
Summary of the collected database (literature database).
Category Input parameter Unit Modelled symbol Range

Minimum Maximum Mean Standard

Deviation

Aggregates Aggregate size range mm x1 2.3 31.5 18.4 6.04

RAC-Replacement % x2 0.00 100 59.71 34.31
Type of RACc -a x3 1 2 NAb NA
Binder Cement typed - x4 1 10 NA NA
Cement content kg/mc x5 155 600 371.11 77.49
Water to cement ratio - x6 0.3 0.66 0.49 0.09
Additives Waste kg/mc x7 1.75 100 26.33 33.20
Nanomaterials kg/mc x8 0.2 1050 178.03 320.59
Fiber kg/mc x9 0.1 127.5 15.29 37.47
Other admixtures kg/mc x10 0.155 194 40.05 57.19
Testing Condition Mixing methode - x11 1 9 NA NA
Output variable Compressive_strength_28_days MPa y 2.2 63 37.78 11.65
a
Unitless variable.
b
Not applicable for categorical variables.
c
RAC: (1) Coarse aggregate, (2) Fine aggregate.
d
Cement type: (1) CEM I/ 42.5R, (2) CEM II/ 42.5R, (3) OPC of Grade 53, (4) OPC of Grade43, (5) CEM-IIAS, (6) CEM-IIIB, (7) Blended Portland, (8) SCR, (9)
geopolymer, (10) OPC and metakaolin.
e
Mixing method: (1) normal mix, (2) ASTM, (3) JGJ/T 221–2010, (4) mixtures possessing, (5) conventional, (6) equivalent mortar, (7) Double DM, (8) Triple TM.

Supervised learning has a collection of examples that specify each input, several components with multi-correlated interactions. Thus, convolu­
output value, or function. The learning system seeks to generate a hy­ tion layers are used in this work to represent the modelling part. The
pothesis that predicts the function or connection between input and developed paradigm attempts to find latent links between output/input
output. However, in unsupervised learning, a collection of learning in­ variables and input variables while maintaining the highest feasible
stances exists in which just the number of inputs is known, but no in­ prediction accuracy. Reference [184] proposed the Residual Network
formation about the proper output is provided. Unsupervised learning (ResNet) idea for the DL paradigm to smooth gradient propagation via
organizes inputs or predicts future values based on the present situation massively deep structures and solve their problems. ResNet residual
[181]. The modeling done in this study is categorized under supervised shortcuts skip several network layers to prevent high-accuracy training,
learning. allowing layers to identify residual functions based on layer inputs
DL (or Deep Neural Network) approaches were created to address the rather than unreferenced ones [185]. We can now define the ResNet
issue of hand-crafted features in sophisticated ML systems [175,176]. notion using DRNNs (i.e., Deep Residua1 Neural Networks).
Thus, DL algorithms may enable both feature connections to desired
outputs and feature extraction [182]. Finally, with sufficient training, • The DRNNs’ depth is set to 36 layers with three unique sections:
the DL system may determine a straight mapping from main or raw • The first input layer segment includes a convolution layer (CL), a
inputs to target outputs without removing features. It may also identify rectified linear unit activation layer (ReLU), and a batch normali­
abstract (high-level) characteristics as a hierarchy that explains basic zation layer (BNL).
(low-level) learned features. This capacity enables DL algorithms to • The second part has a stack of four Residua1 Building Units (ResBU),
decompose difficult tasks into simpler issues and solve them [183]. which serves as the structure’s core.
GRAC compressive strength is a complicated occurrence involving

9
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 3. The framework of analyzing dynamic modulus parameters.

• The final portion completes the construction with three layers: a fully Z
connected layer (FCL), a Softmax layer, and an output layer.
( ) e j̃
f zj̃ = K (5)
∑ Z
ej
The first layer normalizes the input data before feeding it into the j=1
network. This data normalization process filters the incoming data array
by subtracting it from the mean values. Table 5 lists the mathematical 1∑ r t (
∑ )2
symbols utilized in the proposed DRNNs, which have the following Loss = [ ypj − ymj ]r (6)
r i=1 j=1
major functions:
t ( )2
Hl+1 = ReLU[f (Hl ) + id(Hl ) ] (1) ∑
ypj − ymj
j=1
(( )) MSE = (7)
cj − μB t
ĉi = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (2)
(σ 2B + ∈ ) Each ResBU has two divisions: main and residual. The fundamental
division consists of six layers, two of each CL, ReLU, and BNL, with the
zj = γ ĉi + β (3) BNL acting as a mediator between the CL and the ReLU. This layout is
{ consistent throughout the proposed DRNNs design. The identical func­
( ) 0, &zj < 0 tion allows the residual division to skip the first five layers (i.e., the
Relu zj = (4)
zj , &zj ≥ 0 shortcut connection) of the CL-BNL-ReLU-CL-BNL division. Then,
shortly before the final ReLU, the main division gets the residual di­
vision’s output elementwise. Eq. (1) is run throughout all ResBUs cal­
culations. Interestingly, the id(Hl ) has no additional computing load to
smooth the gradient flow in the shortcut link [184,186]. As seen in
Fig. 4-(a), each ResBU has residual and main branches. An activation
layer for BN and ReLU follows each Conv layer. The shortcut link (re­
Table 5
sidual branch) to the identity function of input values eliminates the
Nomenclature.
main branch’s layers. As shown in Fig. 4-(b), the residual branch output
Symbol Definition
is added item by item to the main branch output just before the final
Hl is the input of the lth ResBU ReLU activation layer. Each ResBU calculates Eq. (1). The existence of a
id(Hl ) is the identity function shortcut link to the identity function smoothes gradient flow without
Hl+1 is the output of the lth ResBU
increasing computing complexity [184].
f(Hl ) is the residual function
y is compressive strength after 28 days The CL acts as the name implies: "convolution layer." It convolves the
ym is the mean of the measured compressive strength values input data by sliding filters over the layer. It can acquire competent
yp is the mean of the predicted compressive strength values internal features using these filters (learnable kernels), which extract
ĉi is the normalized activation value
local representations from incoming data. The feature maps are
cj is the batch input value
is the mini-batch variance
computed using a series of sliding filters that combine the dot product of
σ2B
μB is the mini-batch mean their associated weights, inputs, and bias values [187]. The CL’s output
β is a learnable offset is sent to the BNL to develop a robust regularization approach to prevent
ϵ is an Epsilon attribute data overfitting [188]. Fortunately, the BNL speeds up the training
γ is a learnable scale factor process by expanding and reorienting each input channel. Furthermore,
zj is the input values of the Softmax layer
m is a subscript that denotes measured values
it reduces the architecture’s responsiveness to startup circumstances
p is a subscript that denotes predicted values [189]. Eq. (3) represents the BNL’s mini-batch involvement in scaling
n is the number of total data points and adjusting the normalized activation of the input calculated in Eq.
t is the number of testing data points (2). Learnable factors (γ and β) determine the scaling and shifting step.
r is the sequence length
The Epsilon characteristic (ϵ) ensures that the numerical stability of the

10
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 4. The proposed DRNNs architecture.

mini-batch variance value is kept to a minimal throughout network depends on these metrics values. The literature is full of such mea­
training. Feeding the BNL output to the ReLU introduces non-linear and surements, each of which differs in how it assesses model correctness
complicated abstraction learnability into the network [190]. Eq. (4) [196]. This research employs a number of well-known performance
applies a threshold operation inside the ReLU function to each element measures to validate/test the defined prediction problem, such as the
(zj̃ ) in order to avoid negative values. At the same time, the input size coefficient of determination (R2), root mean square error (RMSE), mean
will remain unchanged. The ReLU function was chosen because of its absolute percentage error (MAPE), and residual plots [197]. These sta­
superior performance over other activation functions (such as saturating tistical parameters of fit may be calculated separately for any model. R2
hyperbolic-tang and sigmoid functions). This method allows for faster calculates the correlation between expected and observed values as
and more accurate generalizations without the need for extensive deep follows:
training [191,192]. n (
∑ )2
The neurons in the FCL are coupled to the neurons in the former layer ypi − ymi
in an all-to-all manner. This holistic link sends all previously learned R2 = 1 − i=1n (8)
∑
qualities to the FCL for the most accurate scenic forecast [189,193]. The (ymi − ym )2
i=1
FCL is then linked to the Softmax layer to forecast the likelihood of
compressive strength values. Eq. (5) employs a normalized exponential Because it is misleading to depend just on one metric of goodness-of-
function as the Softmax function for multi-class classification purposes. fit statistics (e.g., R2) for accuracy assessment, the availability of col­
It is considered a generalization of the logistic-sigmoidal function [194]. lective accuracy metrics may have a major impact on making correct
The loss function (Loss) and mean-squared error (MSE) are determined judgments regarding model behavior. As a result, other indicators of
at the output layer (also known as the conclusion layer). The difference global bias should be studied. To assess model bias, a residual plot shows
between the estimated and measured compressive strength is controlled the difference between the observed response and the values of the fitted
using Eq.s (6 and 7). During the training and setup phases, 36 filters of response. The ideal residual plot, also known as the null residual plot,
size four are employed. Each CL has thirty-six feature maps generated by shows a random dispersion of points that results in a band of generally
these filters. Furthermore, a mini-batch of size 256 is employed using the constant width around the identity line. The residual plot is used to
stochastic-gradient-descent approach. This approach is ideal for weight check the model’s fit and assumptions, such as constant variance,
initialization in DL networks, particularly when employing the ReLU normality, and error independence, to verify whether they are correct
layer [195]. [198,199].
The (RMSE) is a quadratic scoring method that determines the
4.2. Goodness of fit statistics average error size. The MAPE is a linear metric that shows that all in­
dividual variances contribute equally to the mean error [200]. The de­
The ML model used for this research is thought to be of interest since parture from the predicted values was assessed using the RMSE and
it would replace the actual laboratory/environment processes for per­ MAPE, which are as follows:
forming various experiments using the specified GSA procedure. √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
Implementing the GRAC compressive strength prediction in a numerical √ n ymi − ypi 2
√ ( y )
√ mi
model would be useful, since the VBSA approach needs several samples RMSE = i=1 × 100, (9)
to converge, see the results section. The significance of goodness of fit n
statistics is highlighted here since the reliability of the VBSA conclusions

11
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

n
∑ impacts induced by the interaction with other factors. Both indices run
ymi − ypi
⌈ ymi
⌉ from 0 to 1, with 1 indicating that the factor is the only regulator of y
MAPE = i=1 × 100, (10) variation and 0 indicating that the factor does not influence the pro­
n
jected values’ uncertainty. As a result, Sfi would help rank input variables
4.3. Variance based GSA [205]. Nonetheless, the low Sfi value is inadequate evidence for the
analyst to deem that variable insignificant. However, STi is the sufficient
Although DL approaches are effective prediction tools, they may be and necessary condition that the input variable does not impact the
regarded as uninformative modeling tools because to the difficulties in output factor. Thus, the total effect index might be used to filter the
evaluating their findings [201]. To remedy this problem, SA may be input variables.
utilized as a compensatory step to determine the cause-and-effect rela­ Analytically, probability density functions should be defined to
tionship between the trained DRNNs model’s outputs and inputs. This provide indices values for the variables being studied. Fortunately, the
part introduces the VBSA for compressive strength prediction. Inte­ extremely precise depiction of the DRNNs in the findings aids in deter­
grating the VBSA into the model analysis allows for a more thorough mining the VBSA with high confidence using LHS simulation, which
evaluation of the correlations between the RAC and input factors. allows for the examination of the whole input factors space [204]. The
Furthermore, it enables the quantitative assessment of each input vari­ VBSA-LHS is summarized in next steps:
able’s contribution to compressive strength. Initially, a sample (T) is created as follows:
To reduce the experimental load, a model-based technique is pro­ ⎛ ⎞
x11 x12 … x1i x1k
posed to infer the local reaction by constructing the partial derivatives ⎜ x21 x22 x2i k ⎟
shown below: T=⎜ ⎝ ⋮
⎟ (14)
⋮ ⋱ ⋮ ⋮ ⎠
Consider the output variable: xμ1 xμ2 … xμi xμk
Compressive strength after 28 days→y = f(x1, x2,…, xi …., xk): Rk→R
∂y where k and μ are the input variable number and sample size, respec­
Si = , ∀ X (11) tively. The sample T is structured in a matrix with, μ rows. Each row
∂xi
represents the outcome of a trial set from the LHS.
where; Si is the index of sensitivity that scales the variable, xi. rele­
vance in establishing y magnitude/level. Unfortunately, the mode-based 1. Split T into two equal samples: A and B.
LSA has two major shortcomings. First, certain calibrated models are
2. Create a sample B matrix by substituting B’s xi vector with corre­
nondifferential. Second, the Si value in Eq. (11) produces deceptive re­
sponding A’s.
sults when the model’s input variables (xi) are significantly correlated.
3. Use the DRNNs model to compute the associated output vectors Y:
In other words, this formulation demands tight assumptions about the
⎛ ⎞
proposed model’s normalcy, linearity, and monotonicity and, thus, is
⎜ y
μ/2
⎛ 1 ⎞
only appropriate for models with closed functions. y ⎜ μ
⎟
⎟
On the other hand, the VBSA was invented to address these issues. It ⎜ y2 + 1 ⎟
⎜ y ⎟ B ⎜
2 ⎟
⎟
Y =⎝
A ⎜ &Y = ⎜ ⎟ (15)
addresses the complete variation range of the explanatory factors by ⋮ ⎠ ⎜ ⋮ ⎟
⎜ ⎟
using a global set of samples, allowing it to account for variable in­ yμ/2 ⎝ ⎠
teractions (joint effect) between numerous input variables. At the same yμ
time, its variance-based form produces a strong sensitivity index
throughout the whole input factor domain [202,203]. It explores the Finally, the Sfi and STi values are calculated numerically for each
influence of each component on the output variable while leaving all variable as follows:
other variables free to change. The complex DRNNS model describes the 2
YA . YB − E(Y A )2
output variable with multiple intercorrelated input variables in this Sfi =
μ
(16)
study. The VBSA may immediately offer a comprehensive insight into Var(YA )
these parameters’ interaction with the output variable. Two main
indices in the VBSA to measure the input factors importance as follows
2
YB . YB + E(Y A )2
STi = 1− (17)
μ
[204]: Var(Y A )
/
Vxi [EX∼i (y xi )] The primary goal of Eqs. (14)–(17) is to average the influence of all
Sfi = , ∀xi (12) input variable values to get the global effect on the response variable. It
V[y]
is worth emphasizing that the first-order sensitivity described in Eq. (16)
/
EX∼i (Vxi [y X∼i ]) is associated with the fluctuation of y in OAT, while all other variables
STi = , ∀xi (13) are held constant. On the contrary, Eq. (17) averages the variation in y
V[y]
when the variable understudy is fixed, and all other variables are varied
where, xi is the ith variable that is under sensitivity analysis, and X∼i in an AAT manner. If the connection is linear and there is no interaction
reflects the variability of all components except the factor under inves­ between the input variables, this will result in a single value for both
tigation. In Eq. (12), the inner expectation represents the fluctuation of indices. This averaging method is critical for the model since mono­
the total output Y mean across changing all variables (i.e., X∼i ) in their tonicity, independence, and linearity are not guaranteed. It should be
domain space while keeping xi constant. Simultaneously, the outer noted that the sample size μ must be adequate to reflect the real vari­
variance is calculated throughout the domain space of the xi variable. ance/covariance in the data, and its value adjustment will also be shown
Eq. (12) calculation represents the first-order effect (Sfi ) of the variable in the next section.
under study, where it gives its main effect to the total value of y
regardless of its interaction with other input variables. Eq. (13) mea­ 5. Numerical study section
sures to how much xi contributes to the output variance where all other
variables X∼i are fixed. This measures the xi total effect (including in­ 5.1. Experimentation results
teractions) in determining the y value. Both indices go from one to zero.
Sfi evaluates the direct ith factor contribution to the variety of y values The test on the GRAC samples has been conducted during two stages:
while STi reveals the direct influence of the ith factor in addition to the 1) The workability and consistency of fresh concrete is measured by

12
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

slump test before the hardening of concrete, according to [ES: 1658–2/ the CWRB usage as fine aggregate and CWCB as coarse aggregate, which
2008 and ECP203–2007], 2) Compressive strength test have been per­ gives a higher bond with the cement paste. The incorporation of CWPB
formed on hardened concrete, see example in Fig. 5.a. The strength of waste reduced the density and bond strength of the GRAC, which
concrete results is depicted in Fig. 6, where all the experimental results decreased with the increasing quantity of CWPB by 20 % of plastic ag­
are summarized in Table 6. The results show that compressive strength gregates. This reduction can be attributed to some particles being
decreases with an increase in w/c ratio. Moreover, an increase in the w/c angular and others having non-uniform shapes, resulting in less fluidity.
ratio indicates an increase in the workability of concrete. Thus, the Apparently, increasing the content of CWPB led to a decrease in
strength of concrete is inversely proportional to the workability of compressive strength, as the same result was reported in [210]. With
concrete. Workability assures that GRAC can be easily handled and has higher quantities of plastic, the mechanical properties decreased.
an impact on its structural applications. The effect of different per­ However, the thermal performance of plastic concrete was seven times
centages of RAC at GRAC mixes on the workability of fresh concrete was better than that of reference concrete. Studies have shown that the
examined using the slump test. The slump of GRAC concrete mixtures incorporation of waste tire rubber aggregates reduces the strength, in­
ranged from 120 mm for the GRAC5 mix for w/c of 0.55–180 mm for the creases permeability, and decreases the thermal conductivity of concrete
GRAC16 mix compared to the much higher for w/c of 0.95. The higher [211]. However, almost all the use of CWRB in different mixes showed a
slump values in the case of 40 % CWPB concrete can be due to lower positive impact, with 0.75 w/c, on energy absorbed. With less content of
water absorption capacity. Other researchers have also reported a CWPB (by 20 %) of the RAC, the amount appeared to be beneficial in the
similar observation [206,207]. This reduction can be attributed to the case of the compression test. The test results showed that high-strength
fact that some particles are angular, and others have non-uniform concrete made with 20 P-40 T-40B-0.75 increased by about 42 %
shapes, resulting in less fluidity. Additionally, mixtures with a w/c compared to that made with dolomite as aggregate for the control mix.
ratio of 0.95 showed more workability than those of 0.55 with an According to the brittle fracture theory, failure initiation in a spec­
amount of 30 %. imen occurs through the largest crack oriented normally to the applied
Density directly affects concrete strength, whereas higher-density load. This principle underscores the stochastic nature of failure, where
components give denser concrete, leading to less void and ultimately factors such as the size and shape of the specimen influence its strength.
increasing strength. Therefore, a strong correlation exists between A larger specimen inherently carries a higher probability of containing
density and compressive strength, which is agreed upon in reference critical cracks that can initiate failure. Table 6 and Fig. 5.b illustrate the
[208]. Almost Lightweight concrete can be classed according to its unit split tensile strength of various mixes. These parameters, including
weight or density, which normally ranges from 320 to 1920 kg/m3, water-cement ratio and recycled aggregate content, play crucial roles in
according to the [ACI Committee 213] guide for structural lightweight determining the energy released at the onset of cracking. In certain
aggregate concrete [209]. In this study, there are different lightweight cases, cracks may be impeded from propagating further due to the
concrete type divisions in terms of strength range, which are low-density presence of large pores in the CWPB or more ductile material in the
concretes (1.70–10.25 MPa), moderate-strength concretes cementitious waste rubber tire (CWRT). Consequently, additional en­
(5.85–20.50 MPa), and structural concretes (4.18–13.75 MPa). The ergy is required to overcome these obstacles, potentially affecting the
density of these concretes is in the range of 1100–1362 kg/m3, overall tensile strength of the concrete [212]. The expected trend sug­
1190–1425 kg/m3, and 1120–1362 kg/m3 with w/c values of 0.55, 0.75 gests a decrease in tensile strength with increasing w/c content. How­
and 0.95, respectively. Furthermore, it can be observed that the 0.75 w/c ever, this effect can be mitigated by the meticulous selection of GRAC
content in the GRAC 5 specimen increased compressive strength by 40 % parameters during concrete production. A relative tensile strength loss
compared to the reference sample with RAC obtained by crushing the of approximately 12.50 % with w/c equal to 0.55 compared to 0.75 and
pre-tested cubes and cylinders. However, workability increased by 30 % 0.95 was observed. Interestingly, the tensile strength values remain
with a density reduction. The increase in compressive strength is due to relatively consistent between w/c of 0.75 and 0.95. These findings

Fig. 5. GRAC compressive and tensile strength tests.

13
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 6. Compressive strength for GRAC mixes.

underscore the importance of understanding the interplay between criteria assessment is done with respect to other ML methods reviewed
material properties and mixture design parameters in GRAC production, in the literature since the collected data is unique to the literature, where
ultimately informing strategies for optimizing tensile strength and the comparison with the literature-reported accuracies is not a fair task.
enhancing the durability of concrete structures. Table 7 validates the DRNNs’ high performance using the metrics dis­
It is worth noting that the use of waste materials like bricks in con­ cussed in the goodness-of-fit statistics section compared to best-reported
crete mixes impacts water absorption, affecting workability and hy­ ML methods. The comparable accuracy results between the testing and
dration. To address this issue, additional water is often added to the mix training stages prove that the model has truly learned the patterns and
to ensure proper hydration and maintain workability. Reference [213] not just overfit the data. Overall, model bias is examined using the re­
found that RA absorbs more water, requiring a 10–15 % increase in the sidual plots in Fig. 9. The model seems to have a normal distribution of
w/c, whereas a 5–20 % increase in water demand for brick aggregates is errors with substantial evidence of error independence. These validation
noted in [214]. Reference [215] confirmed that fine recycled aggregates and testing stages are critical for assuring the fitting of the DRNNS since
need more water to achieve similar workability to conventional con­ they measure how well its architecture understands the links between
crete, ensuring proper hydration and satisfactory mechanical properties. regulating parameters. If a model learned the latent features of the RAC
However, for the sake of modeling accuracy, no additional water is components perfectly and their rule in determining compressive
added beyond the stated w/c. Interestingly, this compensation is ach­ strength, the VBSA might accurately identify the essential input
ieved by the need for higher w/c values, specifically 0.55, 0.75, and parameters.
0.95. Also, in our green mix designs, we mixed recycled fine aggregate In the literature on material properties, few studies have examined
and recycled coarse aggregate in a dry state by hand before adding the transfer learning in ML to reduce the number of needed experimental
required quantity of wa-ter and mixing thoroughly. Additionally, we specimens in the current laboratory [163,164]. Different percentages of
selected the SRC for its very high compressive strength and durability the tested GRAC specimen results are examined to fine-tune the DRNNs
under aggressive conditions. Its very low heat of hydration also helps to parameters in the training process. Fig. 10 reports the accuracy of these
prevent shrinkage cracks. attempts with respect to the R2 values. Sixty percent of the laboratory
data was sufficient to obtain reasonable accuracy, which resulted in a
40 % reduction in laboratory efforts.
5.2. Evaluation of DRNNS

MATLAB 2021a software is used to apply all the methods developed 5.3. VBSA results
in this study using a personal computer with a 2.6 GHz Intel® Xeon®
CPU, 64 GB of RAM, and a GeForce GTX 1050 graphics card. The Following validation, the developed DRNNs model is now ready for
literature database’s outputs and inputs are normalized for the proposed the VBSA analysis setup. To assess k input components divided into φ
DRNNs architecture, and the data set is divided randomly into two levels, resulting in a φk-by-φk grid of points. The analyzed scenario has
groups: training and testing, where 80 % of the database’s data points around eleven trillion prediction points for up to eleven input compo­
are used for training, with the remaining 20 % for testing. nents, with ten values evaluating on average. To determine required
In the parameter setting and training of the DRNNs model, each Conv samples number, two sampling procedures might be used: deterministic
layer generates 30 feature maps using 30 four-size kernels. A stochastic and stochastic. Deterministic techniques use a strategy for selecting μ
gradient descent (SGD) with a mini-batch size of 256 was used. The samples to cover as much of the huge area as possible systematically and
initial learning rate is set to 0.01 for weight initialization, and it reduces uniformly. It makes an effort to cover all areas within the point’s
exponentially to 0.01 after 300 Epochs. As shown in Fig. 7, all of these domain. Stochastic techniques take input data as random variables with
parameters are set during early hyperparameter optimization runs. specified probability density functions (PDFs) and generate the μ sam­
Fig. 8 shows the quality of the DRNNS predictions compared to the ples using these PDFs. This research used the LHS with a deterministic
original data. It proves the high capabilities of the model to track the selection technique applicable to all GSA methods. Furthermore, the
high variance in the compressive strength as the target output. A multi- deterministic selection technique is compatible with the fact that our

14
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Table 6 Table 6 (continued )

Experimental results for the GRAC. Mix NO. Mix. Fresh concrete test Compressive strength Tensile
Mix NO. Mix. Fresh concrete test Compressive strength Tensile (GRAC) Code (MPa) strength
(GRAC) Code (MPa) strength (MPa)
(MPa) Density Slump 7 14 28 28 days
Density Slump 7 14 28 28 days (Kg\m3) (mm) days days days
(Kg\m3) (mm) days days days
GRAC18 40P- 1140 170 1.85 3.75 4.18 1.75
Control RAC 1606 120 4.62 7.10 14.10 1.10 30T-
GRAC1 20P- 1362 120 2.60 3.70 5.00 1.60 30B-
20T- 0.95
60B-
0.55
GRAC2 20 P- 1224 130 4.60 7.50 10.25 2.10
40 T- Table 7
40B- Goodness-of-fit statistics for different ML tools.
0.55
Model Training data Testing data
GRAC3 20P- 1325 150 2.30 2.60 3.50 1.10
2
60T- R RMSE% MAPE% R2 RMSE% MAPE%
20B-
ANN 0.72 15.34 11.56 0.67 18 14.22
0.55
SVM 0.85 7.87 6.43 0.81 9.45 6.73
GRAC4 20P- 1398 160 4.20 7.50 12.45 2.32
DT 0.87 5.45 4.68 0.86 6.67 4.88
20T-
GEP 0.91 4.33 3.42 0.89 4.98 3.60
60B-
DRNNs 0.93 3.98 3.33 0.91 4.26 3.56
0.75
GRAC5 20P- 1356 165 6.80 10.50 20.50 3.00
40T-
40B- input component values are obtained from a design mixing paradigm
0.75 and predetermined environmental circumstances. LHS is a statistical
GRAC6 20P- 1425 170 3.40 6.20 9.25 1.75 method for generating a sample of parameter values from a quasi-
60T- random multivariate distribution. In terms of statistical sampling, a
20B-
square grid containing sample locations is a Latin square if (and only if)
0.75
GRAC7 20P- 1342 165 3.20 5.10 6.50 1.70 there is only one sample in each row and column [216].
20T- A Latin hypercube is an extension of this concept to an infinite
60B- number of dimensions, with each sample unique inside each axis-
0.95
aligned hyperplane containing it. When sampling a function with n
GRAC8 20P- 1298 165 4.50 8.70 13.75 2.50
40T- variables, the range of each variable is partitioned into φ intervals with
40B- equal probability [217]. The sample points are then positioned to meet
0.95 the LHS requirements, ensuring that the number of divisions for each
GRAC9 20P- 1298 170 2.80 5.60 6.20 1.75 variable is equal. This independence is one of the main advantages of
60T-
this sampling strategy. Another advantage is that random samples may
20B-
0.95 be gathered one at a time while maintaining track of which samples have
GRAC10 40P- 1085 130 1.76 2.50 3.32 1.05 already been taken. In the LHS, one must first decide how many sample
20T- points to use and then recollect the row and column where each sample
40B-
point was taken [218].
0.55
GRAC11 40P- 1100 135 3.25 5.07 6.85 2.30
The robustness of assessing VBSA indices increases with sample size
40T- μ, but is limited by available processing resources. As a result, the
20B- question of the appropriate number of model evaluations necessary to
0.55 construct valid sensitivity indices is often discussed. Robustness, in this
GRAC12 40P- 1152 150 1.50 1.70 2.35 0.85
instance, means that the VBSA index value does not change much as μ
30T-
30B- increases. μ is intuitively controlled by the number of variables and their
0.55 domains. Nonetheless, the selected VBSA technique requires a certain
GRAC13 40P- 1230 170 2.80 5.20 8.30 1.55 number of tests to verify that its indices are stable. The μ selection is
20T-
studied in Fig. 11, in which the relationship between the fluctuation of
40B-
0.75
sensitivity indices and μ is depicted. A sample size of 13,000 is sufficient
GRAC14 40P- 1190 170 4.62 7.10 13.65 2.30 for the main effect and total effect indices to converge.
30T- With the predefined sample size, the analytical methodology pre­
30B- f
sented in subsection 4.3 is used to estimate the main effects (VB (Si ))
0.75
GRAC15 40P- 1257 170 2.30 4.20 5.85 1.23 and total effects (VB (STi ))
indices, which are reported in Fig. 12. The
30T- VBSA recommends that cement content be the most influential param­
30B-
eter in terms of its first and total effects. Waste content comes in second
0.75
GRAC16 40P- 1120 180 2.15 3.80 5.23 1.70
place with respect to the first-order index, reflecting the high direct
20T- sensitivity of the RAC compressive strength to the waste over all other
40B- additives. Although w/c achieved a moderate main effect, it ranked
0.95 second in the total effects, reflecting two logical properties: first, there is
GRAC17 40P- 1125 170 3.67 6.36 9.25 2.50
a direct interaction between w/c and cement content, and second, there
40T-
20B- is indirect interaction with all other mixture components. Unsurpris­
0.95 ingly, the VBSA method’s use of two indices (first order and total order)
allowed it to capture the entangled impact of the contributing compo­
nents. All studied factors showed significant effects except the mixing

15
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 7. DRNNs hyperparameters optimization process.

Fig. 8. True values vs. DRNNS prediction values.

method, which achieved low values for both indices. elements in concrete mixes have received a lot of attention. Using pre­
dictive models for concrete qualities may save time and resources while
6. Conclusions providing information on scheduling tasks such as framework removal.
This study introduces a powerful DL technique as a new solution
Concrete is the most widely utilized man-made material in struc­ methodology. The DL’s superior performance originates from its ca­
tures, pavements, and dams. Concrete needs a huge amount of fine and pacity to acquire more highly complicated features (i.e., characteristics)
coarse materials. To protect natural resources, waste and byproduct from raw data than other ML algorithms. The suggested depth is assessed

16
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 9. Residuals (measured values – predicted values) plot for different data points.

Fig. 10. Pre-trained DRNNS prediction values against different GRAC training data percentages.

by the increasing number of layers, which often improves feature input parameter’s contribution to the anticipated compressive strength
extraction capabilities. The intelligence of DL architectures is shown by value. The VBSA aided with LHS is introduced for this rule. Deriving
coping with the vanishing gradients issue, which arises with a very deep stable VBSA indices required 13,000 data points, which could not have
network. The vanishing gradients issue leads to substantial training er­ been achieved without the DRNNs model. The validation of the DRNNs
rors and considerable accuracy deterioration. The proposed DRNNs with other ML tools gives confidence about the obtained VBSA results.
model is introduced as an innovative structure of the DL that has the Cement and waste contents, in addition to w/c, were the most influential
potential to learn more complicated and latent relationships of the RAC factors, whereas mixing type was the least factor. This research topic
mixture components than ordinary DL. Different GRAC mixtures are holds significant importance in the construction industry due to its focus
proposed to investigate the possibility of using RA from various sources, on advancing environmental sustainability, resource efficiency, and
such as plastic and rubber waste aggregates and demolition waste, as a innovation in concrete technology. By exploring the utilization of the
prospective solution. The database for training the DRNNs is composed RAC and incorporating waste materials like tires, plastic, and red brick
of both experimentation done in this study and data collected from the into concrete mixes, the study aims to reduce environmental impact and
literature. The pre-trained DRNNs model managed to transfer the promote resource conservation. The development of accurate models for
practice of other labs to the current one, reducing the required efforts by predicting mechanical strength, such as the proposed DRNNs, offers cost
40 % while attaining reasonable accuracy. Like other advanced ML ap­ and time savings by streamlining material evaluation processes. Addi­
proaches, the DRNNs mode lacks obvious formulae for analyzing each tionally, the research provides valuable insights for decision-makers by

17
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

Fig. 11. First order (main effect) and total order (total effect) sensitivity analysis convergence diagrams.

f
Fig. 12. The variance-based sensitivity indices values of main effects (VB (Si )) and total effects (VB (STi )).

identifying influential mixture components on compressive strength, Declaration of Competing Interest

aiding in informed material selection and construction practices. Over­
all, this research contributes to sustainable construction practices and “On behalf of all authors, the corresponding author states that there
fosters innovation in the concrete industry, aligning with sustainability is no conflict of interest”
goals and promoting efficient, greener construction materials and pro­
cesses. The proposed framework can potentially pave the way for new Data availability
concrete mixture research where more comprehensive datasets and new
GSA tools can further derive new insights for the studied mechanical Data will be made available on request.
properties and other similar ones.
References
CRediT authorship contribution statement
[1] D.-C. Feng, Z.-T. Liu, X.-D. Wang, Y. Chen, J.-Q. Chang, D.-F. Wei, Z.-M. Jiang,
Machine learning-based compressive strength prediction for concrete: an
Mahmoud Owais: Conceptualization, Methodology, statistical adaptive boosting approach, Constr. Build. Mater. 230 (2020) 117000.
interpretation, Modeling, Validation, Writing - Original draft. Lamiaa [2] M.M. Moein, A. Saradar, K. Rahmati, S.H.G. Mousavinejad, J. Bristow,
K.Idriss: Formal analysis, Data collection, Lab procedures, Experimen­ V. Aramali, M. Karakouzian, Predictive models for concrete properties using
machine learning and deep learning approaches: a review, J. Build. Eng. 63
tation, Writing - review & editing. (2023) 105444.

18
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

[3] M. Nili, V. Afroughsabet, Combined effect of silica fume and steel fibers on the microstructure of recycled aggregate concrete, Constr. Build. Mater. 24 (5) (2010)
impact resistance and mechanical properties of concrete, Int. J. Impact Eng. 37 701–708.
(8) (2010) 879–886. [37] B. Lu, C. Shi, Z. Cao, M. Guo, J. Zheng, Effect of carbonated coarse recycled
[4] M. Nataraja, N. Dhang, A. Gupta, Statistical variations in impact resistance of concrete aggregate on the properties and microstructure of recycled concrete,
steel fiber-reinforced concrete subjected to drop weight test, Cem. Concr. Res. 29 J. Clean. Prod. 233 (2019) 421–428.
(7) (1999) 989–995. [38] L.K. Idriss, Y.A.S. Gamal, Properties of rubberized concrete prepared from
[5] D.-C. Feng, J. Li, Stochastic nonlinear behavior of reinforced concrete frames. II: different cement types, Recycling 7 (3) (2022) 39.
Numerical simulation, J. Struct. Eng. 142 (3) (2016) 04015163. [39] A. Cotto-Ramos, S. Dávila, W. Torres-García, A. Cáceres-Fernández, Experimental
[6] D. Feng, X. Ren, J. Li, Stochastic damage hysteretic model for concrete based on design of concrete mixtures using recycled plastic, fly ash, and silica
micromechanical approach, Int. J. Non-Linear Mech. 83 (2016) 15–25. nanoparticles, Constr. Build. Mater. 254 (2020) 119207.
[7] B.A. Young, A. Hall, L. Pilon, P. Gupta, G. Sant, Can the compressive strength of [40] M.J. Miah, R. Huaping, S.C. Paul, A.J. Babafemi, R. Sharma, J.G. Jang,
concrete be estimated from knowledge of the mixture proportions?: New insights Performance of eco-friendly concrete made from recycled waste tire fine
from statistical analysis and machine learning methods, Cem. Concr. Res. 115 aggregate as a replacement for river sand, in: Structures, Elsevier, 2023.
(2019) 379–388. [41] S. Raffoul, R. Garcia, D. Escolano-Margarit, M. Guadagnini, I. Hajirasouliha,
[8] S. Mindess, Developments in the Formulation and Reinforcement of Concrete, K. Pilakoutas, Behaviour of unconfined and FRP-confined rubberised concrete in
Woodhead Publishing, 2019. axial compression, Constr. Build. Mater. 147 (2017) 388–397.
[9] S.H. Kosmatka, W.C. Panarese, B. Kerkhoff, Design and control of concrete [42] İ.B. Topçu, A. Unverdi, Scrap tires/crumb rubber, Waste and Supplementary
mixtures, Portland Cement Association Skokie, IL, 2002. Cementitious Materials in Concrete, Elsevier, 2018, pp. 51-77.
[10] P. Bischoff, S. Perry, Compressive behaviour of concrete at high strain rates, [43] S. Frondistou-Yannas, Waste concrete as aggregate for new concrete, J. Proc.
Mater. Struct. 24 (1991) 425–450. (1977) 373–376.
[11] H.-s Shi, B.-w Xu, X.-c Zhou, Influence of mineral admixtures on compressive [44] S. Frondistou-Yannas, T. Itoh, Economic feasibility of concrete recycling,
strength, gas permeability and carbonation of high performance concrete, Constr. J. Struct. Div. 103 (4) (1977) 885–899.
Build. Mater. 23 (5) (2009) 1980–1985. [45] A. Ajdukiewicz, A. Kliszczewicz, Influence of recycled aggregates on mechanical
[12] M.J. Munir, S.M.S. Kazmi, Y.-F. Wu, X. Lin, M.R. Ahmad, Development of a novel properties of HS/HPC, Cem. Concr. Compos. 24 (2) (2002) 269–279.
compressive strength design equation for natural and recycled aggregate concrete [46] C. Européen, Eurocode 2: Design of concrete structures—part 1-1: General rules
through advanced computational modeling, J. Build. Eng. 55 (2022) 104690. and rules for buildings, London: British Standard Institution (2004).
[13] M. Fm Zain, S.M. Abd, Multiple regression model for compressive strength [47] M.C, CEB-FIP, Model code for concrete structures, Bull. D. ’Inf. 516 (1990).
prediction of high performance concrete, J. Appl. Sci. 9 (1) (2009) 155–160. [48] A. Committee, Building code requirements for structural concrete (ACI 318-08)
[14] B. Bharatkumar, R. Narayanan, B. Raghuprasad, D. Ramachandramurthy, Mix and commentary, American Concrete Institute, 2008.
proportioning of high performance concrete, Cem. Concr. Compos. 23 (1) (2001) [49] K. Rahal, Mechanical properties of concrete with recycled coarse aggregate,
71–80. Build. Environ. 42 (1) (2007) 407–415.
[15] S. Bhanja, B. Sengupta, Investigations on the compressive strength of silica fume [50] B. Wang, L. Yan, Q. Fu, B. Kasal, A comprehensive review on recycled aggregate
concrete using statistical methods, Cem. Concr. Res. 32 (9) (2002) 1391–1394. and recycled aggregate concrete, Resour., Conserv. Recycl. 171 (2021) 105565.
[16] R. Bekkerman, M. Bilenko, J. Langford, Scaling up machine learning: Parallel and [51] J. Xiao, W. Li, Y. Fan, X. Huang, An overview of study on recycled aggregate
distributed approaches, Cambridge University Press, 2011. concrete in China (1996–2011), Constr. Build. Mater. 31 (2012) 364–383.
[17] M. Kanevski, V. Timonin, A. Pozdnukhov, Machine learning for spatial [52] G. Dimitriou, P. Savva, M.F. Petrou, Enhancing mechanical and durability
environmental data: theory, applications, and software, EPFL press, 2009. properties of recycled aggregate concrete, Constr. Build. Mater. 158 (2018)
[18] K.P. Murphy, Machine learning: a probabilistic perspective, MIT press, 2012. 228–235.
[19] S. Marsland, Machine learning: an algorithmic perspective, Chapman and Hall/ [53] C. Shi, Z. Wu, Z. Cao, T.C. Ling, J. Zheng, Performance of mortar prepared with
CRC, 2011. recycled concrete aggregate enhanced by CO2 and pozzolan slurry, Cem. Concr.
[20] B. Morris, The components of the wired spanning forest are recurrent, Probab. Compos. 86 (2018) 130–138.
Theory Relat. Fields 125 (2) (2003) 259–265. [54] A.R. Anjam, H. Faghihmaleki, Laboratory evaluation of the performance of
[21] X. Jiang, S. Mahadevan, H. Adeli, Bayesian wavelet packet denoising for recycled aggregate concrete containing construction and stone factories waste in
structural system identification, Structural Control and Health Monitoring: The terms of compressive and tensile strength, Math. Probl. Eng. 2020 (2020) 1–8.
Official Journal of the International Association for Structural Control and [55] W. Bai, Y. Geng, C. Yuan, J. Guan, Y. Lv, C. Xie, J. Lei, Study on mechanical
Monitoring and of the European Association for the Control of Structures 14(2) properties and mesoscopic damage mechanism of metakaolin modified recycled
(2007) 333-356. aggregate concrete, Dev. Built Environ. 17 (2024) 100332.
[22] L.K. Idriss, M. Owais, Global sensitivity analysis for seismic performance of shear [56] X. Hu, L. Ben, Z. Chen, Z. Yu, L. Kai-hang, Z. Can-hao, A new mesoscopic
wall with high-strength steel bars and recycled aggregate concrete, Constr. Build. calculation model of chloride ion erosion in recycled coarse aggregate concrete
Mater. 411 (2024) 134498. (RAC): characteristic fractal dimension of pore structure and service life
[23] O. Abdeljaber, O. Avci, D.J. Inman, Active vibration control of flexible cantilever prediction, Dev. Built Environ. 16 (2023) 100282.
plates using piezoelectric materials and artificial neural networks, J. Sound Vib. [57] C. Vintimilla, M. Etxeberria, Z. Li, Durable Structural Concrete Produced with
363 (2016) 33–53. Coarse and Fine Recycled Aggregates Using Different Cement Types,
[24] A. Behnood, K.P. Verian, M.M. Gharehveran, Evaluation of the splitting tensile Sustainability 15 (19) (2023) 14272.
strength in plain and steel fiber-reinforced concrete based on the compressive [58] C. Liang, H. Ma, Y. Pan, Z. Ma, Z. Duan, Z. He, Chloride permeability and the
strength, Constr. Build. Mater. 98 (2015) 519–529. caused steel corrosion in the concrete with carbonated recycled aggregate,
[25] S.M. Levy, P. Helene, Durability of recycled aggregates concrete: a safe way to Constr. Build. Mater. 218 (2019) 506–518.
sustainable development, Cem. Concr. Res. 34 (11) (2004) 1975–1980. [59] S. Yang, H. Lee, Drying shrinkage and rapid chloride penetration resistance of
[26] F. Olorunsogo, N. Padayachee, Performance of recycled aggregate concrete recycled aggregate concretes using cement paste dissociation agent, Materials 14
monitored by durability indexes, Cem. Concr. Res. 32 (2) (2002) 179–185. (6) (2021) 1478.
[27] M. Malešev, V. Radonjanin, S. Marinković, Recycled concrete as aggregate for [60] M.G. Beltrán, A. Barbudo, F. Agrela, A.P. Galvín, J.R. Jiménez, Effect of cement
structural concrete production, Sustainability 2 (5) (2010) 1204–1225. addition on the properties of recycled concretes to reach control concretes
[28] Y. Yan, Y. Si, Y. Zheng, X. Wang, Durability of nano-reinforced recycled aggregate strengths, J. Clean. Prod. 79 (2014) 124–133.
concrete under load and chloride ingress, Materials 15 (18) (2022) 6194. [61] H.M. Allujami, M. Abdulkareem, T.M. Jassam, R.A. Al-Mansob, A. Ibrahim, J.
[29] K.M. Elhadi, T. Ali, M.Z. Qureshi, N. Anwar, O. Zaid, A. Majdi, M. Qaisar, L. Ng, H.C. Yam, Mechanical properties of concrete containing recycle concrete
Improving the engineering properties of sustainable recycled aggregate concrete aggregates and multi-walled carbon nanotubes under static and dynamic stresses,
modified with metakaolin, Case Stud. Constr. Mater. 19 (2023) e02430. Case Stud. Constr. Mater. 17 (2022) e01651.
[30] J. Thomas, N.N. Thaickavil, P. Wilson, Strength and durability of concrete [62] P. Rajhans, S.K. Panda, S. Nayak, Sustainable self compacting concrete from C&D
containing recycled concrete aggregates, J. Build. Eng. 19 (2018) 349–365. waste by improving the microstructures of concrete ITZ, Constr. Build. Mater. 163
[31] P. Soroushian, Strength and durability of recycled aggregate concrete containing (2018) 557–570.
milled glass as partial replacement for cement, Constr. Build. Mater. 29 (2012) [63] P. Nuaklong, A. Wongsa, K. Boonserm, C. Ngohpok, P. Jongvivatsakul, V. Sata,
368–377. P. Sukontasukkul, P. Chindaprasirt, Enhancement of mechanical properties of fly
[32] J. Xie, H. Zhang, L. Duan, Y. Yang, J. Yan, D. Shan, X. Liu, J. Pang, Y. Chen, X. Li, ash geopolymer containing fine recycled concrete aggregate with micro carbon
Effect of nano metakaolin on compressive strength of recycled concrete, Constr. fiber, J. Build. Eng. 41 (2021) 102403.
Build. Mater. 256 (2020) 119393. [64] G. Lin, K. Liu, Y. Chen, Y. Ji, R. Jiang, Experimental Investigation on Compressive
[33] Y.-W. Choi, D.-J. Moon, J.-S. Chung, S.-K. Cho, Effects of waste PET bottles Properties of Fiber Recycled Aggregate Concrete, J. Renew. Mater. 11 (11)
aggregate on the properties of concrete, Cem. Concr. Res. 35 (4) (2005) 776–781. (2023).
[34] M. Nili, V. Afroughsabet, The effects of silica fume and polypropylene fibers on [65] D. Lu, H. Cao, Q. Shen, Y. Gong, C. Zhao, X. Yan, Dynamic characteristics and
the impact resistance and mechanical properties of concrete, Constr. Build. Mater. chloride resistance of basalt and polypropylene fibers reinforced recycled
24 (6) (2010) 927–933. aggregate concrete, Adv. Polym. Technol. 2020 (2020) 1–9.
[35] P. Matar, J.J. Assaad, Concurrent effects of recycled aggregates and [66] Z. Liu, C. Cai, H. Peng, F. Fan, Experimental study of the geopolymeric recycled
polypropylene fibers on workability and key strength properties of self- aggregate concrete, J. Mater. Civ. Eng. 28 (9) (2016) 04016077.
consolidating concrete, Constr. Build. Mater. 199 (2019) 492–500. [67] Z. Zheng, P. Deng, Mechanical and fracture properties of slag/steel slag-based
[36] D. Kong, T. Lei, J. Zheng, C. Ma, J. Jiang, J. Jiang, Effect and mechanism of geopolymer fully recycled aggregate concrete, Constr. Build. Mater. 413 (2024)
surface-coating pozzalanics materials around aggregate on properties and ITZ 134533.

19
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

[68] A. Ababneh, M. Alhassan, M. Abu-Haifa, Predicting the contribution of recycled [95] F. Deng, Y. He, S. Zhou, Y. Yu, H. Cheng, X. Wu, Compressive strength prediction
aggregate concrete to the shear capacity of beams without transverse of recycled concrete based on deep learning, Constr. Build. Mater. 175 (2018)
reinforcement using artificial neural networks, Case Stud. Constr. Mater. 13 562–569.
(2020) e00414. [96] C.-Y. Fan, P.-C. Chang, J.-J. Lin, J. Hsieh, A hybrid model combining case-based
[69] A. Ahmad, G. Kotsovou, D.M. Cotsovos, N.D. Lagaros, Assessing the accuracy of reasoning and fuzzy decision tree for medical data classification, Appl. Soft
RC design code predictions through the use of artificial neural networks, Int. J. Comput. 11 (1) (2011) 632–644.
Adv. Struct. Eng. 10 (2018) 349–365. [97] K.-Y. Tung, C. Huang, S.-L. Chen, C.-T. Shih, Mining the Generation Xers’ job
[70] A. Ahmad, V. Plevris, Q.-u-Z. Khan, Prediction of properties of FRP-confined attitudes by artificial neural network and decision tree—empirical evidence in
concrete cylinders based on artificial neural networks, Crystals 10 (9) (2020) 811. Taiwan, Expert Syst. Appl. 29 (4) (2005) 783–794.
[71] F. Khademi, M. Akbari, S.M. Jamal, M. Nikoo, Multiple linear regression, artificial [98] Z. Yu, F. Haghighat, B.C. Fung, H. Yoshino, A decision tree method for building
neural network, and fuzzy logic prediction of 28 days compressive strength of energy demand modeling, Energy Build. 42 (10) (2010) 1637–1646.
concrete, Front. Struct. Civ. Eng. 11 (2017) 90–99. [99] A. Gholampour, I. Mansouri, O. Kisi, T. Ozbakkaloglu, Evaluation of mechanical
[72] A.T.A. Dantas, M.B. Leite, K. de Jesus Nagahama, Prediction of compressive properties of concretes containing coarse recycled concrete aggregates using
strength of concrete containing construction and demolition waste using artificial multivariate adaptive regression splines (MARS), M5 model tree (M5Tree), and
neural networks, Constr. Build. Mater. 38 (2013) 717–722. least squares support vector regression (LSSVR) models, Neural Comput. Appl. 32
[73] Z.-H. Duan, S.-C. Kou, C.-S. Poon, Using artificial neural networks for predicting (2020) 295–308.
the elastic modulus of recycled aggregate concrete, Constr. Build. Mater. 44 [100] A. Behnood, J. Olek, M.A. Glinicki, Predicting modulus elasticity of recycled
(2013) 524–532. aggregate concrete using M5′ model tree algorithm, Constr. Build. Mater. 94
[74] H. Naderpour, O. Poursaeidi, M. Ahmadi, Shear resistance prediction of concrete (2015) 137–147.
beams reinforced by FRP bars using artificial neural networks, Measurement 126 [101] A. Behnood, E.M. Golafshani, Machine learning study of the mechanical
(2018) 299–308. properties of concretes containing waste foundry sand, Constr. Build. Mater. 243
[75] J. Xu, X. Zhao, Y. Yu, T. Xie, G. Yang, J. Xue, Parametric sensitivity analysis and (2020) 118152.
modelling of mechanical properties of normal-and high-strength recycled [102] A. Behnood, V. Behnood, M.M. Gharehveran, K.E. Alyamac, Prediction of the
aggregate concrete using grey theory, multiple nonlinear regression and artificial compressive strength of normal and high-performance concretes using M5P
neural networks, Constr. Build. Mater. 211 (2019) 479–491. model tree algorithm, Constr. Build. Mater. 142 (2017) 199–207.
[76] N. Deshpande, S. Londhe, S. Kulkarni, Modeling compressive strength of recycled [103] Y. Ayaz, A.F. Kocamaz, M.B. Karakoç, Modeling of compressive strength and UPV
aggregate concrete by artificial neural network, model tree and non-linear of high-volume mineral-admixtured concrete using rule-based M5 rule and tree
regression, Int. J. Sustain. Built Environ. 3 (2) (2014) 187–198. model M5P classifiers, Constr. Build. Mater. 94 (2015) 235–240.
[77] R. Siddique, P. Aggarwal, Y. Aggarwal, Prediction of compressive strength of self- [104] J.-S. Chou, C.-K. Chiu, M. Farfoura, I. Al-Taharwa, Optimizing the prediction
compacting concrete containing bottom ash using artificial neural networks, Adv. accuracy of concrete compressive strength based on a comparison of data-mining
Eng. Softw. 42 (10) (2011) 780–786. techniques, J. Comput. Civ. Eng. 25 (3) (2011) 242–253.
[78] Z.-H. Duan, S.-C. Kou, C.-S. Poon, Prediction of compressive strength of recycled [105] S. Mangalathu, J.-S. Jeon, Classification of failure mode and prediction of shear
aggregate concrete using artificial neural networks, Constr. Build. Mater. 40 strength for reinforced concrete beam-column joints using machine learning
(2013) 1200–1206. techniques, Eng. Struct. 160 (2018) 85–94.
[79] C. Deepa, K. SathiyaKumari, V.P. Sudha, Prediction of the compressive strength of [106] Q. Han, C. Gui, J. Xu, G. Lacidogna, A generalized method to predict the
high performance concrete mix using tree based modeling, Int. J. Comput. Appl. 6 compressive strength of high-performance concrete by improved random forest
(5) (2010) 18–24. algorithm, Constr. Build. Mater. 226 (2019) 734–742.
[80] M. Nikoo, F. Torabian Moghadam, Ł. Sadowski, Prediction of concrete [107] J. Zhang, G. Ma, Y. Huang, F. Aslani, B. Nener, Modelling uniaxial compressive
compressive strength by evolutionary artificial neural networks, Adv. Mater. Sci. strength of lightweight self-compacting concrete using random forest regression,
Eng. 2015 (2015). Constr. Build. Mater. 210 (2019) 713–719.
[81] A.A. Shahmansouri, M. Yazdani, S. Ghanbari, H.A. Bengar, A. Jafari, H.F. Ghatte, [108] M. Jalal, H. Jalal, RETRACTED: Behavior assessment, regression analysis and
Artificial neural network model to predict the compressive strength of eco- support vector machine (SVM) modeling of waste tire rubberized concrete,
friendly geopolymer concrete incorporating silica fume and natural zeolite, Elsevier (2020).
J. Clean. Prod. 279 (2021) 123697. [109] B.G. Aiyer, D. Kim, N. Karingattikkal, P. Samui, P.R. Rao, Prediction of
[82] H. Naderpour, A.H. Rafiean, P. Fakharian, Compressive strength prediction of compressive strength of self-compacting concrete using least square support
environmentally friendly concrete using artificial neural networks, J. Build. Eng. vector machine and relevance vector machine, KSCE J. Civ. Eng. 18 (2014)
16 (2018) 213–219. 1753–1758.
[83] J. Amani, R. Moeini, Prediction of shear strength of reinforced concrete beams [110] M.R. Kaloop, A.R. Gabr, S.M. El-Badawy, A. Arisha, S. Shwally, J.W. Hu,
using adaptive neuro-fuzzy inference system and artificial neural network, Sci. Predicting resilient modulus of recycled concrete and clay masonry blends for
Iran. 19 (2) (2012) 242–248. pavement applications using soft computing techniques, Front. Struct. Civ. Eng.
[84] S. Kumar, S. Barai, Neural networks modeling of shear strength of SFRC corbels 13 (2019) 1379–1392.
without stirrups, Appl. Soft Comput. 10 (1) (2010) 135–148. [111] P.R. Prem, A. Thirumalaiselvi, M. Verma, Applied linear and nonlinear statistical
[85] F. Altun, Ö. Kişi, K. Aydin, Predicting the compressive strength of steel fiber models for evaluating strength of Geopolymer concrete, Comput. Concr. 24 (1)
added lightweight concrete using neural network, Comput. Mater. Sci. 42 (2) (2019) 7–17.
(2008) 259–265. [112] A. Nazari, J.G. Sanjayan, Modelling of compressive strength of geopolymer paste,
[86] R. Perera, M. Barchín, A. Arteaga, A. De Diego, Prediction of the ultimate strength mortar and concrete by optimized support vector machine, Ceram. Int. 41 (9)
of reinforced concrete beams FRP-strengthened in shear using neural networks, (2015) 12164–12177.
Compos. Part B: Eng. 41 (4) (2010) 287–298. [113] K. Yaswanth, V. Sathish Kumar, J. Revathy, G. Murali, C. Pavithra, Compressive
[87] H. Tanarslan, M. Secer, A. Kumanlioglu, An approach for estimating the capacity strength prediction of ternary blended geopolymer concrete using artificial neural
of RC beams strengthened in shear with FRP reinforcements using artificial neural networks and support vector regression, Innov. Infrastruct. Solut. 9 (2) (2024) 32.
networks, Constr. Build. Mater. 30 (2012) 556–568. [114] H.U. Ahmed, R.R. Mostafa, A. Mohammed, P. Sihag, A. Qadir, Support vector
[88] U. Atici, Prediction of the strength of mineral admixture concrete using regression (SVR) and grey wolf optimization (GWO) to predict the compressive
multivariable regression analysis and an artificial neural network, Expert Syst. strength of GGBFS-based geopolymer concrete, Neural Comput. Appl. 35 (3)
Appl. 38 (8) (2011) 9609–9618. (2023) 2909–2926.
[89] A. Behnood, E.M. Golafshani, Predicting the compressive strength of silica fume [115] M.-Y. Cheng, J.-S. Chou, A.F. Roy, Y.-W. Wu, High-performance concrete
concrete using hybrid artificial neural network with multi-objective grey wolves, compressive strength prediction using time-weighted evolutionary fuzzy support
J. Clean. Prod. 202 (2018) 54–64. vector machines inference model, Autom. Constr. 28 (2012) 106–115.
[90] E.M. Golafshani, A. Behnood, M. Arashpour, Predicting the compressive strength [116] S. Motamedi, S. Shamshirband, R. Hashim, D. Petković, C. Roy, RETRACTED:
of normal and High-Performance Concretes using ANN and ANFIS hybridized Estimating unconfined compressive strength of cockle shell–cement–sand
with Grey Wolf Optimizer, Constr. Build. Mater. 232 (2020) 117266. mixtures using soft computing methodologies, Elsevier (2015).
[91] D.-K. Bui, T. Nguyen, J.-S. Chou, H. Nguyen-Xuan, T.D. Ngo, A modified firefly [117] A. Slowik, H. Kwasnicka, Evolutionary algorithms and their applications to
algorithm-artificial neural network expert system for predicting compressive and engineering problems, Neural Comput. Appl. 32 (2020) 12363–12379.
tensile strength of high-performance concrete, Constr. Build. Mater. 180 (2018) [118] A.E. Eiben, S.K. Smit, Parameter tuning for configuring and analyzing
320–333. evolutionary algorithms, Swarm Evolut. Comput. 1 (1) (2011) 19–31.
[92] F. Özcan, C.D. Atiş, O. Karahan, E. Uncuoğlu, H. Tanyildizi, Comparison of [119] C.A.C. Coello, Evolutionary algorithms for solving multi-objective problems,
artificial neural network and fuzzy logic models for prediction of long-term Springer, 2007.
compressive strength of silica fume concrete, Adv. Eng. Softw. 40 (9) (2009) [120] M. Owais, M.K. Osman, Complete hierarchical multi-objective genetic algorithm
856–863. for transit network design problem, Expert Syst. Appl. 114 (2018) 143–154.
[93] D.V. Dao, H.-B. Ly, S.H. Trinh, T.-T. Le, B.T. Pham, Artificial intelligence [121] D.B. Fogel, Evolutionary algorithms in engineering applications [book reviews,
approaches for prediction of compressive strength of geopolymer concrete, IEEE Trans. Evolut. Comput. 2 (2) (1998) 72-72.
Materials 12 (6) (2019) 983. [122] D. Greiner, J. Periaux, D. Quagliarella, J. Magalhaes-Mendes, B. Galván,
[94] İ.B. Topçu, M. Sarıdemir, Prediction of mechanical properties of recycled Evolutionary algorithms and metaheuristics: applications in engineering design
aggregate concretes containing silica fume using artificial neural networks and and optimization, Math. Probl. Eng. 2018 (2018) 1–4.
fuzzy logic, Comput. Mater. Sci. 42 (1) (2008) 74–82. [123] M. Owais, M.K. Osman, G. Moussa, Multi-objective transit route network design
as set covering problem, IEEE Trans. Intell. Transp. Syst. 17 (3) (2015) 670–679.

20
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

[124] M. Baioletti, G. Di Bari, A. Milani, V. Poggioni, Differential evolution for neural [155] A. Sieber, S. Uhlenbrook, Sensitivity analyses of a distributed catchment model to
networks optimization, Mathematics 8 (1) (2020) 69. verify the model structure, J. Hydrol. 310 (1-4) (2005) 216–235.
[125] H.Q. Nguyen, H.-B. Ly, V.Q. Tran, T.-A. Nguyen, T.-T. Le, B.T. Pham, [156] M.-J. Shin, J.H. Guillaume, B.F. Croke, A.J. Jakeman, Addressing ten questions
Optimization of artificial intelligence system by evolutionary algorithm for about conceptual rainfall–runoff models with global sensitivity analyses in R,
prediction of axial capacity of rectangular concrete filled steel tubes under J. Hydrol. 503 (2013) 135–152.
compression, Materials 13 (5) (2020) 1205. [157] J. Nossent, P. Elsen, W. Bauwens, Sobol’sensitivity analysis of a complex
[126] Y. Sun, Y. Gao, A multi-objective particle swarm optimization algorithm based on environmental model, Environ. Model. Softw. 26 (12) (2011) 1515–1525.
gaussian mutation and an improved learning strategy, Mathematics 7 (2) (2019) [158] T.G. Nguyen, J.-L. de Kok, Systematic testing of an integrated systems model for
148. coastal zone management using sensitivity and uncertainty analyses, Environ.
[127] M. Owais, Traffic sensor location problem: Three decades of research, Expert Syst. Model. Softw. 22 (11) (2007) 1572–1587.
Appl. 208 (2022) 118134. [159] J. Dang, J. Zhao, S. Dai Pang, S. Zhao, Durability and microstructural properties
[128] G. Abdollahzadeh, E. Jahani, Z. Kashir, Predicting of compressive strength of of concrete with recycled brick as fine aggregates, Constr. Build. Mater. 262
recycled aggregate concrete by genetic programming, Comput. Concr. 18 (2) (2020) 120032.
(2016) 155–163. [160] Z. Xiao, T.-C. Ling, C.-S. Poon, S.-C. Kou, Q. Wang, R. Huang, Properties of
[129] M.F. Iqbal, Q.-f Liu, I. Azim, X. Zhu, J. Yang, M.F. Javed, M. Rauf, Prediction of partition wall blocks prepared with high percentages of recycled clay brick after
mechanical properties of green concrete incorporating waste foundry sand based exposure to elevated temperatures, Constr. Build. Mater. 49 (2013) 56–61.
on gene expression programming, J. Hazard. Mater. 384 (2020) 121322. [161] L. Xu, F. Wu, Y. Chi, P. Cheng, Y. Zeng, Q. Chen, Effects of coarse aggregate and
[130] I.F. Kara, Prediction of shear strength of FRP-reinforced concrete beams without steel fibre contents on mechanical properties of high performance concrete,
stirrups based on genetic programming, Adv. Eng. Softw. 42 (6) (2011) 295–304. Constr. Build. Mater. 206 (2019) 97–110.
[131] A.H. Gandomi, A.H. Alavi, S. Kazemi, M. Gandomi, Formulation of shear strength [162] M. Gopinath, P. Abimaniu, C.D. Rishi, K. Pravinkumar, P. Tejeshwar,
of slender RC beams using gene expression programming, part I: Without shear Experimental investigation on waste plastic fibre concrete with partial
reinforcement, Autom. Constr. 42 (2014) 112–121. replacement of coarse aggregate by recycled coarse aggregate, Mater. Today:
[132] A.H. Gandomi, A.H. Alavi, M. Gandomi, S. Kazemi, Formulation of shear strength Proc. (2023).
of slender RC beams using gene expression programming, part II: with shear [163] G.S. Moussa, M. Owais, Modeling Hot-Mix asphalt dynamic modulus using deep
reinforcement, Measurement 95 (2017) 367–376. residual neural Networks: Parametric and sensitivity analysis study, Constr. Build.
[133] A.H. Gandomi, A.H. Alavi, M.G. Sahab, New formulation for compressive strength Mater. 294 (2021) 123589.
of CFRP confined concrete cylinders using linear genetic programming, Mater. [164] G.S. Moussa, M. Owais, Pre-trained deep learning for hot-mix asphalt dynamic
Struct. 43 (2010) 963–983. modulus prediction with laboratory effort reduction, Constr. Build. Mater. 265
[134] A.A. Shahmansouri, H.A. Bengar, E. Jahani, Predicting compressive strength and (2020) 120239.
electrical resistivity of eco-friendly concrete containing natural zeolite via GEP [165] M. Owais, G.S. Moussa, Global sensitivity analysis for studying hot-mix asphalt
algorithm, Constr. Build. Mater. 229 (2019) 116883. dynamic modulus parameters, Constr. Build. Mater. 413 (2024) 134775.
[135] S.M. Mousavi, P. Aminian, A.H. Gandomi, A.H. Alavi, H. Bolandi, A new [166] M. Owais, Analysing Witczak 1-37A, Witczak 1-40D and Modified Hirsch Models
predictive model for compressive strength of HPC using gene expression for asphalt dynamic modulus prediction using global sensitivity analysis, Int. J.
programming, Adv. Eng. Softw. 45 (1) (2012) 105–114. Pavement Eng. 24 (1) (2023) 2268808.
[136] S.B. Aval, H. Ketabdari, S.A. Gharebaghi, Estimating shear strength of short [167] G.S. Moussa, M. Owais, E. Dabbour, Variance-based global sensitivity analysis for
rectangular reinforced concrete columns using nonlinear regression and gene rear-end crash investigation using deep learning, Accid. Anal. Prev. 165 (2022)
expression programming, Structures, Elsevier, 2017, pp. 13-23. 106514.
[137] M. Sarıdemir, Genetic programming approach for prediction of compressive [168] B. En, 197-1: 2011, Cement part 1 (2011) 7-43.
strength of concretes containing rice husk ash, Constr. Build. Mater. 24 (10) [169] D.P. Bentz, Three-dimensional computer simulation of Portland cement hydration
(2010) 1911–1919. and microstructure development, J. Am. Ceram. Soc. 80 (1) (1997) 3–21.
[138] E.M. Golafshani, A. Ashour, Prediction of self-compacting concrete elastic [170] P.K. Mehta, P. Monteiro, Concrete: microstructure, properties, and materials, (No
modulus using two symbolic regression techniques, Autom. Constr. 64 (2016) Title) (2006).
7–19. [171] C. Rosental, Certifying knowledge: The sociology of a logical theorem in artificial
[139] P.O. Awoyera, M.S. Kirgiz, A. Viloria, D. Ovallos-Gazabon, Estimating strength intelligence, Am. Sociol. Rev. (2003) 623–644.
properties of geopolymer self-compacting concrete using machine learning [172] P. Langley, The changing science of machine learning, Mach. Learn. 82 (3) (2011)
techniques, J. Mater. Res. Technol. 9 (4) (2020) 9016–9028. 275–279.
[140] B.A. Omran, Q. Chen, R. Jin, Comparison of data mining techniques for predicting [173] A.C. Müller, S. Guido, Introduction to machine learning with Python: a guide for
compressive strength of environmentally friendly concrete, J. Comput. Civ. Eng. data scientists, " O’Reilly Media, Inc."2016.
30 (6) (2016) 04016029. [174] T.G. Dietterich, Ensemble methods in machine learning, International workshop
[141] E.M. Golafshani, A. Behnood, Automatic regression methods for formulation of on multiple classifier systems, Springer, 2000, pp. 1-15.
elastic modulus of recycled aggregate concrete, Appl. Soft Comput. 64 (2018) [175] M. Haenlein, A. Kaplan, A brief history of artificial intelligence: On the past,
377–400. present, and future of artificial intelligence, Calif. Manag. Rev. 61 (4) (2019)
[142] Z.M. Yaseen, R.C. Deo, A. Hilal, A.M. Abd, L.C. Bueno, S. Salcedo-Sanz, M. 5–14.
L. Nehdi, Predicting compressive strength of lightweight foamed concrete using [176] O. Avci, O. Abdeljaber, S. Kiranyaz, M. Hussein, M. Gabbouj, D.J. Inman,
extreme learning machine model, Adv. Eng. Softw. 115 (2018) 112–125. A review of vibration-based damage detection in civil structures: from traditional
[143] J.-S. Chou, C.-F. Tsai, A.-D. Pham, Y.-H. Lu, Machine learning in concrete strength methods to Machine Learning and Deep Learning applications, Mech. Syst. Signal
simulations: Multi-nation data analytics, Constr. Build. Mater. 73 (2014) Process. 147 (2021) 107077.
771–780. [177] A. Marani, M.L. Nehdi, Machine learning prediction of compressive strength for
[144] I. Nunez, A. Marani, M.L. Nehdi, Mixture optimization of recycled aggregate phase change materials integrated cementitious composites, Constr. Build. Mater.
concrete using hybrid machine learning model, Materials 13 (19) (2020) 4331. 265 (2020) 120286.
[145] M. Owais, A.I. Shahin, Exact and heuristics algorithms for screen line problem in [178] E. Alpaydin, Introduction to machine learning, MIT press, 2020.
large size networks: shortest path-based column generation approach, IEEE Trans. [179] M. Ivanovic, M. Radovanovic, Modern machine learning techniques and their
Intell. Transp. Syst. 23 (12) (2022) 24829–24840. applications, Int. Conf. Electron., Commun. Netw. (2015).
[146] M. Owais, A.S. Ahmed, G.S. Moussa, A.A. Khalil, Design scheme of multiple- [180] W.Z. Taffese, E. Sistonen, Machine learning for durability and service-life
subway lines for minimizing passengers transfers in mega-cities transit networks, assessment of reinforced concrete structures: recent advances and future
Int. J. Rail Transp. 9 (6) (2021) 540–563. directions, Autom. Constr. 77 (2017) 1–14.
[147] M. Owais, A. Alshehri, J. Gyani, M.H. Aljarbou, S. Alsulamy, Prioritizing rear-end [181] W.Z. Taffese, E. Sistonen, J. Puttonen, CaPrM: Carbonation prediction model for
crash explanatory factors for injury severity level using deep learning and global reinforced concrete using machine learning methods, Constr. Build. Mater. 100
sensitivity analysis, Expert Syst. Appl. 245 (2024) 123114. (2015) 70–82.
[148] C.W. Schwartz, R. Li, H. Ceylan, S. Kim, K. Gopalakrishnan, Global sensitivity [182] J. Schmidhuber, Deep learning in neural networks: an overview, Neural Netw. 61
analysis of mechanistic–empirical performance predictions for flexible (2015) 85–117.
pavements, Transp. Res. Rec. 2368 (1) (2013) 12–23. [183] M. Owais, Deep learning for integrated origin–destination estimation and traffic
[149] M.D. Morris, Factorial sampling plans for preliminary computational sensor location problems, IEEE Trans. Intell. Transp. Syst. (2024).
experiments, Technometrics 33 (2) (1991) 161–174. [184] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition,
[150] A. Saltelli, M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D. Gatelli, M. Saisana, Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (2016) 770–778.
S. Tarantola, Global sensitivity analysis: the primer, John Wiley & Sons, 2008. [185] A. Alshehri, M. Owais, J. Gyani, M.H. Aljarbou, S. Alsulamy, Residual neural
[151] R. Spear, G. Hornberger, Eutrophication in peel inlet—II. Identification of critical networks for origin–destination trip matrix estimation from traffic sensor
uncertainties via generalized sensitivity analysis, Water Res. 14 (1) (1980) 43–49. information, Sustainability 15 (13) (2023) 9881.
[152] F. Pianosi, K. Beven, J. Freer, J.W. Hall, J. Rougier, D.B. Stephenson, T. Wagener, [186] M. Owais, A.S. Ahmed, G.S. Moussa, A.A. Khalil, An optimal metro design for
Sensitivity analysis of environmental models: a systematic review with practical transit networks in existing square cities based on non-demand criterion,
workflow, Environ. Model. Softw. 79 (2016) 214–232. Sustainability 12 (22) (2020) 9566.
[153] M. Owais, A.S. Ahmed, Frequency based transit assignment models: Graph [187] J. Zang, L. Wang, Z. Liu, Q. Zhang, G. Hua, N. Zheng, Attention-based temporal
formulation study, IEEE Access 10 (2022) 62991–63003. weighted convolutional neural network for action recognition. IFIP International
[154] S. Tarantola, W. Becker, D. Zeitz, A comparison of two sampling methods for Conference on Artificial Intelligence Applications and Innovations, Springer,
global sensitivity analysis, Comput. Phys. Commun. 183 (5) (2012) 1061–1072. 2018, pp. 97–108.

21
M. Owais and L.K. Idriss Construction and Building Materials 440 (2024) 137393

[188] S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network training by [205] M. Owais, G. Moussa, Y. Abbas, M. El-Shabrawy, Optimal frequency setting for
reducing internal covariate shift, arXiv preprint arXiv:1502.03167 (2015). circular bus routes in urban areas, JES J. Eng. Sci. 41 (5) (2013) 1796–1811.
[189] I. Goodfelow, Y. Bengio, A. Courville, Deep Learning (Adaptive Computation and [206] R.V. Silva, J. de Brito, N. Saikia, Influence of curing conditions on the durability-
Machine Learning Series), e MIT Press, Cambridge, England, 2016. related performance of concrete made with selected plastic waste aggregates,
[190] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, Cem. Concr. Compos. 35 (1) (2013) 23–31.
V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, Proc. IEEE Conf. [207] Y. Ghernouti, B. Rabehi, B. Safi, R. Chaid, Use of recycled plastic bag waste in the
Comput. Vis. Pattern Recognit. (2015) 1–9. concrete, J. Int. Sci. Publ.: Mater., Methods Technol. 8 (2011) 480–487.
[191] P. Ramachandran, B. Zoph, Q.V. Le, arXiv preprint, arXiv:1710.05941 (, Search. [208] R. Othman, R.P. Jaya, K. Muthusamy, M. Sulaiman, Y. Duraisamy, M.M.A.
Act. Funct. (2017). B. Abdullah, A. Przybył, W. Sochacki, T. Skrzypczak, P. Vizureanu, Relation
[192] M. Owais, G.S. Moussa, K.F. Hussain, Robust deep learning architecture for traffic between density and compressive strength of foamed concrete, Materials 14 (11)
flow estimation from a subset of link sensors, J. Transp. Eng., Part A: Syst. 146 (1) (2021) 2967.
(2020) 04019055. [209] H. Prayuda, F. Asani, M.B. Santoso, F. Wirawan, Influence of expanded clay
[193] M. Owais, G.S. Moussa, K.F. Hussain, Sensor location model for O/D estimation: aggregate on the engineering properties of lightweight concrete, Ing. ía e Invest.
multi-criteria meta-heuristics approach, Oper. Res. Perspect. 6 (2019) 100100. óN. 44 (1) (2024) e106174.
[194] C.M. Bishop, Pattern recognition and machine learning, springer, 2006. [210] M. Záleská, M. Pavlíková, J. Pokorný, O. Jankovský, Z. Pavlík, R. Černý,
[195] K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: surpassing human- Structural, mechanical and hygrothermal properties of lightweight concrete based
level performance on imagenet classification, Proc. IEEE Int. Conf. Comput. Vis. on the application of waste plastics, Constr. Build. Mater. 180 (2018) 1–11.
(2015) 1026–1034. [211] Y.C. Khern, S.C. Paul, S.Y. Kong, A.J. Babafemi, V. Anggraini, M.J. Miah,
[196] M. Owais, A. Alshehri, Pareto optimal path generation algorithm in stochastic B. Šavija, Impact of chemically treated waste rubber tire aggregates on
transportation networks, IEEE Access 8 (2020) 58970–58981. mechanical, durability and thermal properties of concrete, Front. Mater. 7 (2020)
[197] T.K. Pellinen, Investigation of the use of dynamic modulus as an indicator of hot- 90.
mix asphalt peformance, (2002). [212] R. Silva, J. De Brito, R. Dhir, Tensile strength behaviour of recycled aggregate
[198] C.-L. Tsai, Z. Cai, X. Wu, The examination of residual plots, Stat. Sin. (1998) concrete, Constr. Build. Mater. 83 (2015) 108–118.
445–465. [213] C.S. Poon, Z. Shui, L. Lam, Effect of microstructure of ITZ on compressive strength
[199] Z. Hussain, A. Akbar, Diagnostics through residual plots in binomial regression of concrete prepared with recycled aggregates, Constr. Build. Mater. 18 (6)
addressing chemical species data, Math. Probl. Eng. 2022 (2022). (2004) 461–468.
[200] M. Owais, A.S. Ahmed, G.S. Moussa, A.A. Khalil, Integrating underground line [214] F. Debieb, S. Kenai, The use of coarse and fine crushed bricks as aggregate in
design with existing public transportation systems to increase transit network concrete, Constr. Build. Mater. 22 (5) (2008) 886–893.
connectivity: Case study in Greater Cairo, Expert Syst. Appl. 167 (2021) 114183. [215] L. Evangelista, J. De Brito, Mechanical behaviour of concrete made with fine
[201] L. Jiang, Y. Xie, X. Wen, T. Ren, Modeling highly imbalanced crash severity data recycled concrete aggregates, Cem. Concr. Compos. 29 (5) (2007) 397–401.
by ensemble methods and global sensitivity analysis, J. Transp. Saf. Secur. (2020) [216] F. Zhang, L. Cheng, M. Wu, X. Xu, P. Wang, Z. Liu, Performance analysis of two-
1–23. stage thermoelectric generator model based on Latin hypercube sampling, Energy
[202] M. Saisana, A. Saltelli, S. Tarantola, Uncertainty and sensitivity analysis Convers. Manag. 221 (2020) 113159.
techniques as tools for the quality assessment of composite indicators, J. R. Stat. [217] A. Karolczuk, M. Kurek, Fatigue life uncertainty prediction using the Monte Carlo
Soc.: Ser. A (Stat. Soc. ) 168 (2) (2005) 307–323. and Latin hypercube sampling techniques under uniaxial and multiaxial cyclic
[203] M. Owais, A.E. Matouk, A factorization scheme for observability analysis in loading, Int. J. Fatigue 160 (2022) 106867.
transportation networks, Expert Syst. Appl. 174 (2021) 114727. [218] E. Bulut, E.I. Albak, G. Sevilgen, F. Öztürk, A new approach for battery thermal
[204] A. Saltelli, M. Ratto, S. Tarantola, F. Campolongo, E. Commission, Sensitivity management system design based on Grey Relational Analysis and Latin
analysis practices: strategies for model-based inference, Reliab. Eng. Syst. Saf. 91 Hypercube Sampling, Case Stud. Therm. Eng. 28 (2021) 101452.
(10-11) (2006) 1109–1125.

22