Journal of Building Engineering 71 (2023) 106508

Contents lists available at ScienceDirect

Journal of Building Engineering

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

Intelligent mix design of recycled brick aggregate concrete based

on swarm intelligence
Shiqi Wang a, Peng Xia a, Zhao Wang b, Tao Meng a, Fuyuan Gong a, *
a
College of Civil Engineering and Architecture, Zhejiang University, Hangzhou, China
b
Institute of Urban Innovation, Yokohama National University, Yokohama, Japan

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

Keywords: The mechanical properties of recycled brick aggregate concrete (RBAC) are significantly affected
Recycled brick aggregate concrete by design parameters such as recycled brick aggregate (RBA) replacement ratio and water-cement
Mix ratio design ratio, which not only increases the difficulty of designing RBAC by experimental or theoretical
Machine learning methods, but also makes the mechanical properties (compressive strength, flexural strength, etc.)
Swarm intelligence algorithms of RBAC difficult to match the design requirements. Therefore, an intelligent mix design method
Multi-objective optimization
for RBAC is proposed in this paper, which can determine the best RBAC mix proportion design
scheme according to the design parameters of RBAC (cement content, water-cement ratio and
crushed brick ratio, etc.). This method is compared with the performance of 8 machine learning
algorithms by creating a database containing 8 input variables. The most suitable model for
predicting compressive strength of RBAC is GWO-BP. The predicted/experimental values ob­
tained by the model on the training set and the testing set are 0.98 and 0.96, respectively, and the
determination coefficients R2 are 0.97 and 0.99, respectively. SHapley Additive exPlanations
(SHAP) is used to analyze the mechanism of design parameters on the compressive strength of
RBAC, and the key parameters affecting the compressive strength of RBAC are cement content
(Cement) and water-cement ratio (W/C). The Multi-objective optimization (MOO) model is used
to target the compressive strength, carbon emission and cost of RBAC, and the scheme with the
highest score is determined as the best design scheme of RBAC mix ratio by determining the
Pareto optimal solution set of RBAC. The results show that the design method proposed in this
paper can effectively guide the mix design of RBAC, the machine learning model optimized by
swarm intelligence can effectively predict the compressive strength of RBAC concrete, while the
developed MOO model can effectively determine the optimal mix design method.

Nomenclature

A coefficient vectors of GWO

b the bias of SVM
c penalty coefficient c
c1 acceleration coefficients of PSO
c2 acceleration coefficients of PSO
C penalty parameter of SVM

* Corresponding author.
E-mail address: gongfy@zju.edu.cn (F. Gong).

https://doi.org/10.1016/j.jobe.2023.106508
Received 24 November 2022; Received in revised form 11 March 2023; Accepted 7 April 2023
Available online 7 April 2023
2352-7102/© 2023 Elsevier Ltd. All rights reserved.
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

C1 coefficient vectors of GWO

Cc carbon emission factors of Cement
Cw carbon emission factors of Water
CCB(0-4) carbon emission factors of CB(0–4)
CCB(4-16) carbon emission factors of CB(4–16)
CCT(0-4) carbon emission factors of CT(0–4)
CCT(4-16) carbon emission factors of CT(4–16)
CNA(0-4) carbon emission factors of NA(0–4)
CNA(4-16) carbon emission factors of NA(4–16)
di distances of the solution from the positive-ideal solution
di+ distances of the solution from the negative-ideal solution
D distance between wolf and prey
f(x) predicted value of SVM
fi(x) predicted value of the ith sample
fm(x) the objective function
f(x)mean predicted mean value
Fi the ith solution
Fideal the positive-ideal solution
Fnon-ideal the negative-ideal solution
F(x) the total function
g gamma
gi(x) the functions of inequality constraint
gbestid best position of the swarm
h upper bounds of xj
hi(x) the functions of equation constraint
K the number of trees
l lower bounds of xj
Mc unit prices of Cement
Mw unit prices of Water
MCB(0-4) unit prices of CB(0–4)
MCB(4-16) unit prices of CB(4–16)
MCT(0-4) unit prices of CT(0–4)
MCT(4-16) unit prices of CT(4–16)
MNA(0-4) unit prices of NA(0–4)
MNA(4-16) unit prices of NA(4–16)
MSE mean square error
n number of samples
N dataset
p numbers of inequality constraints
pbestid best position of the single particle
q Structure of XGB
r1i random numbers in the interval [0,1]
r2i random numbers in the interval [0,1]
R correlation coefficient
R2 coefficient of determination
RMSE root mean square error
SD standard deviation
t number of iterations
u numbers of equation constraints
U1 coefficient vectors of GWO
Uc quality of Cement
Uw quality of Water
UCB(0-4) quality of CB(0–4)
UCB(4-16) quality of CB(4–16)
UCT(0-4) quality of CT(0–4)
UCT(4-16) quality of CT(4–16)
UNA(0-4) quality of NA(0–4)
vt velocities of particle i at the tth iterations
vt+1 velocities of particle i at the (t+1)th iterations
w weight vector of SVM
w1 inertia weight of PSO

2
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

x input value of layers of BPNN

xi the input value
xj variables of objective function
xt position of particle i at the tth iterations
xt+1 position of particle i at the (t+1)th iterations
X(t) position of the wolf at moment t
X(t+1) position of the wolf at moment (t+1)
Xp(t) position of the prey at moment t
y output value of layers of BPNN
yi the output value
y(t) predicted value of the round i model of XGB
ε allowable error
ξ* slack variables of SVM
σ smoothness factor
φ(x) feature function
ω the leaf weight of XGB

1. Introduction
With the acceleration of construction industrialization, the massive exploitation of natural sand and low utilization rate of con­
struction waste have become key problems [1–3]. Replacing natural aggregates (NA) with recycled aggregates (RA) not only improves
the utilization of construction waste, but also reduces carbon emissions from concrete materials [4]. Compared with NA, RA has higher
water absorption [5–8], less density [9–11], and larger porosity [12,13], resulting in lower compressive strength [14,15] and flexural
strength [16–18] of recycled aggregate concrete (RAC) compared to those of natural aggregate concrete (NAC). Currently, there is no
uniform design standard for the mix proportion of recycled concrete [19,20]. Due to the differences in properties between RA and NA

Fig. 1. Intelligent mix proportion model design.

3
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

[21], RAC cannot be directly designed according to the mix design method of natural concrete [22]. Usually, RAC should be reasonably
designed to meet the rational requirements [23]. Several experimental studies have been dedicated to the optimal design of recycled
concrete [24,25]. The results show that the amount of recycled aggregate, cement and water cement ratio have a significant effect on
the compressive strength and workability of recycled concrete [26], and optimizing the quality of different mixed strength concrete
aggregate will be beneficial to improve the compressive strength and splitting strength of RAC [27]. In addition, the minimum paste
theory [28] and aggregate skeleton theory [29] are also used to guide the mix design of recycled concrete. But theoretical calculation
methods often have problems such as poor applicability and complex procedures.
Machine learning (ML) has shown excellent performance in predicting concrete performance [30–32]. Scholars have extensively
investigated static performance and durability of concrete using the ML model [33–36], while few studies have been reported on the
use of machine learning models to analyze the mix design of concrete. PSO-BP [37], least squares support vector regression (LSSVR)
[38] and BAS-SVR [39] algorithms are applied to analyze the mix design method of plastic concrete, lightweight foamed concrete and
electrically conductive cementitious composites (ECCC), respectively. Based on the prediction results of ANN algorithm, Šipoš et al.
[40] suggested a mix design method applicable only to the analysis of the compressive strength of RBAC.

Fig. 2. Principles of machine learning algorithms.

4
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

In summary, experiments have the problems of large cost and time consuming, while the calculation procedure of the theoretical
method has poor applicability. The existing research mainly focus on the accuracy of ML model to predict the compressive strength of
recycled brick aggregate concrete and ignore the influence of design parameters such as water-cement proportion on performance of
RBAC, which cannot effectively guide the mix design of RBAC. Therefore, the purpose of this paper is not to analyze the accuracy and
possibility of the ML model to predict the mechanical properties of recycled brick aggregate concrete. Instead, it is expected to propose
an intelligent mix design method for recycled brick aggregate concrete based on multiple perspectives (carbon emissions, cost and
mechanical properties) so that to (1) solve the problems of poor applicability of mix design methods caused by differences in aggregate
performance; (2) utilizing the full potential of concrete materials; (3) provide a basis for the application of recycled brick aggregate
concrete components. Therefore, the main problems studied in this paper are listed as following (specific analysis process is shown in
(Fig. 1)):
● The optimization algorithm is used to optimize the structural performance of the ML model, and to determine the model that is most
suitable for the prediction of RBAC compressive strength.
● The SHAP method is used to eliminate the black box problem of machine learning, study and analyze the influence of several
parameters such as water-cement proportion and aggregate type on the compressive strength of RBAC, and finally determine the
key parameters.
● An intelligent mix design method of RBAC is proposed by considering the compressive strength, carbon emission and cost.

Specifically, the RBAC mix proportion database is established based on the existing test results, 6 typical ML models and 2 ML
models based on Swarm Intelligence (SI) are used to explore the nonlinear relationship between input parameters and compressive
strength of RBAC, which is ignored by traditional design methods, the performance (R2, R, SD, RMSE) of ML model is evaluated by
Taylor diagram, and Shapley Additives Explanations (SHAP) is used to quantify the influence of each input variable on the compressive
strength of RBAC and determine the key influencing parameters; the RBAC mix proportion design model is established considering
compressive strength, carbon emission and cost. Then the Pareto front and optimal solution set of the model are determined. The
Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) method is used to evaluate the optimization results and
determine the optimal mix design scheme of RBAC.

2. Machine learning and Swarm intelligence optimization algorithms

In this study, 8 machine learning models that have been widely used in the field of civil engineering are adopted, and two of them
use the SI algorithm to optimize their network parameters. In this section, six ML models that are not optimized by SI algorithm are
introduced, and the basic structure and characteristics of the models are analyzed. Coefficient of determination (R2), correlation
coefficient (R), standard deviation (SD) and root mean square error (RMSE) are also described to evaluate the performance of ML
model.

2.1. Machine learning algorithms

2.1.1. Back propagation neural network (BPNN)
BPNN is a multi-layer feedforward neural network using error back propagation (Fig. 2 (a)), which consists of an input layer, an
output layer and several hidden layers [41]. The network calculates the error by forward propagation of information and adjusts the
weights and thresholds by error back propagation. Then the network parameters corresponding to the minimum error are determined.
Therefore, the network can effectively deal with the complex problems of multiple variables. It should be noted that the tansig(x) and
purelin(x) functions are used as the activation functions of the input layer-hidden layer and the hidden layer-output layer respectively,
and the specific formulas of two functions are shown as follows：
2
y = tansing(x) = 2x
− 1 (1)
1 + e−

y = purelin(x) = x (2)

where y is the output value of layers of BPNN; x is the input value of layers of BPNN.

2.1.2. Support vector machine (SVM)

SVM has been widely used in regression problems in engineering field because of its strong learning ability and good generalization
ability [42,43]. This method minimizes the sum of distances from all data to that regression plane by mapping the feature vector of
each data to a high-dimensional space and searching for the best regression plane [44]. The given training data set is known as: N =
{(x1, y1), (x2, y2),⋅⋅⋅, (xn, yn)} [25], where xi is the input value and yi is the output value. The regression function of the algorithm can be
expressed as following:
f (x) = wφ(x) + b (3)

where φ(x) is the feature function, w is the weight vector, and b is the bias. The error between the output value yi and the predicted
value f(x) can be analyzed according to the structural risk minimization principle [45], which can be used to analyze this type of
problem by Eq. (4):

5
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

1 ∑n
R(w) = ‖w‖2 + l (yi , f (x)) (4)
2 i=1

Some errors are often allowed in the predicting process, so slack variables ξi and ξ* are introduced to analyze the allowable error ε,
and the regression problem is transformed into an optimization problem, and the corresponding objective function and constraints can
be found in Eq. (5). This type of problem can be solved using Lagrange multipliers [46].

1 ∑n
( )
min R(w) = ‖w‖2 + C ξi + ξ*
2 i=1
⎧
⎪
⎪
⎪
⎪ yi − w · φ(x) − b ≤ ε + ξi
⎪
⎪
⎪
⎪ (5)
⎨ w · φ(x) − yi + b ≤ ε + ξ*
⎪
s.t.
⎪
⎪
⎪ ξ* ≥ 0
⎪
⎪
⎪
⎪
⎪
⎪
⎩ ξi ≥ 0

where C is penalty parameter.

2.1.3. Random forest (RF)

Random forest obtains a new sub-training set by randomly selecting m sample data points from the dataset N [47]. The generated
sub-training set is used for training and multiple regression trees are obtained, where the predicted value obtained by each regression
tree is the mean of all leaf nodes of the sample point, and the predicted value of the final model is the mean of the predicted results of all
regression trees [48].

2.1.4. Extreme learning machine (ELM)

ELM is a feedforward neural network with a single hidden layer [49], and similar to artificial neural networks, this network consist
of an input layer, a hidden layer and an output layer. Each layer is connected by a feature mapping function, and the input value is
processed through the hidden layer and passed to the output layer, and then the output layer derives the predicted value according to
the mapping function [50]. The difference between ELM and BPNN is that ELM can solve for the weight through the error function,
while BPNN updates the weights by back-propagating errors [51].

2.1.5. Extreme gradient boosting (XGB)

XGB is an integrated algorithm based on the idea of Boosting (integrating weak classifiers into strong classifiers), which splits new
trees for prediction by fitting and merging the trees generated in the previous round. Finally the predicted values of all trees are
weighted and summed to obtain the final prediction value [52,53]. The integration model of this algorithm is shown in Eq. (6).
∑
K
y(t) = fk (x) (6)
k=1

where y(t) is the predicted value of the model; K is the number of trees; fk is the quantitative relationship between the structure q
corresponding to the kth tree and the leaf weight ω.

2.1.6. Generalized regression neural network (GRNN)

GRNN accomplishes the prediction of the target value by solving the density function using Parzen nonparametric estimation [54].
The network consists of four parts: input layer, hidden layer, summation layer, and output layer [55]. Compared with traditional
neural networks, GRNN has stronger nonlinear mapping ability and learning ability, and the parameters that GRNN needs to adjust are
much less than those of the traditional network. Moreover, the selection of network structure is dependent on the characteristics of
data samples, which greatly reduces the influence of human intervention, but also requires data sample analysis to reduce the
interference of data noise [56–58].

2.1.7. Evaluation of model performance

In this study, coefficient of determination (R2), correlation coefficient (R), standard deviation (SD) and root mean square error
(RMSE) are used to evaluate the performance of ML model, the corresponding mathematical model are as follows:
( )
(yi − ymean ) fi (x) − f (x)mean
R2 = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (7)
∑n 2 ∑n ( )2
i=1 (yi − ymean ) i=1 fi (x) − f (x)mean

∑n ( )
i=1 (yi − ymean ) fi (x) − f (x)mean
R = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n ̅ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n ( )2 (8)
2
i=1 (yi − ymean ) i=1 fi (x) − f (x)mean

6
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ n
RMSE = (yi − fi (x))2 (9)
n i=1

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1 ∑ n
( ( ) )2
SD = (yi − fi (x) ) − (yi − fi (x) )mean (10)
n − 1 i=1

Where ymean is the mean of the output values; fi(x) is the predicted value of the ith sample obtained by ML model; f(x)mean is the
predicted mean value obtained by ML model; n is the number of samples; where R determines the correlation of the data by calculating
the covariance of the output value and the predicted value and their respective variances; RMSE eliminates the influence of dimension,
making it convenient to compare the performance of different ML models; SD reflects the degree of dispersion of the output value from
the predicted value.

2.2. Swarm intelligence optimization algorithm

As an excellent meta-heuristic algorithm [59], Swarm intelligence optimization algorithm has attracted the attention of scholars,
this kind of algorithm searches for global and local optimal solutions by generating initial population and using heuristic rules [60]. As
typical SI algorithms, PSO and GWO have been widely used in engineering optimization problems [61–63]. Therefore, this paper
proposed to use PSO and GWO to optimize the hyperparameters of ML model (Fig. 3).

2.2.1. Particle swarm optimization (PSO)

The principle of the PSO algorithm is derived from the foraging behavior of birds [64], where the birds find the optimal strategy by
sharing information about the population to enable the group to find the optimal strategy (Fig. 4 (a)). The algorithm guarantees the
optimization capabilities by continuously updating the position and velocity information of a single particle and the population (Eq.
(11)-Eq. (12)). The information of a single particle is determined by analyzing its own historical best position and other particle
positions, and the information of the swarm is guaranteed by finding the information of the optimal particle. The optimization search
process will end when the optimal policy satisfies the constraints such as the number of iterations or the minimum error, the
pseudo-code for this algorithm is shown in Tabel.1a.
( ) ( )
vt+1 = w · vt + c1 · r1i · pbestid − xt + c2 · r2i · gbestid − xt (11)

xt+1 = xt + vt+1 (12)

where vt+1 and vt id are the velocities of particle i at the tth and (t+1)th iterations; xt+1 and xt are the position of particle i at the tth and

Fig. 3. SI optimization of ML model parameters.

7
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

Fig. 4. The optimization process of PSO and GWO.

(t+1)th iterations; w1 denotes inertia weight used to control the convergence ability of the algorithm; pbestid and gbestid are the best
position of the single particle and the best position of the.
swarm, respectively; c1 and c2 are acceleration coefficients; r1i and r2i are two random numbers in the interval [0,1].

2.2.2. Gray wolf optimizer (GWO)

Compared with other meta-heuristic algorithms, Gray Wolf optimizer (GWO) has the advantages of simple structure, few pa­
rameters and self-adaptive adjustment [65], which can effectively guarantee the balance between local optimization and global
optimization. The gray wolf population follows a strict hierarchy (Fig. 4 (b)): there are all gray wolves that can be dominated by wolf α,
the wolf β obeys the wolf α, while the wolf δ obeys the wolf α and wolf β, and the remaining gray wolves need to obey the above 3 Gy
wolves. Using multiple gray wolves for joint search can effectively avoid local optimization stagnation and improve the global optimal
convergence ability [66,67]. In the process of optimization, the α wolf’s information is considered the optimal strategy, the β wolf’s
information is considered the inferior strategy, the δ wolf’s information is the third-ranking strategy, and the remaining gray wolves
seek superiority based on the information provided by the first three wolves (the pseudo-code of this algorithm is shown in Table .1
(b).), the mathematical model of optimization can be expressed as:
⃒ ⃒
D = ⃒C1 − Xp (t) − X(t)⃒ (13)

X(t + 1) = Xp (t) − D · A (14)

where C1 and A are coefficient vectors; D denotes the distance between the wolf and the prey; Xp(t) and X(t) are the position of the prey
and the position of the wolf at moment t, respectively; X(t+1) is the position of the wolf at moment (t+1); t denotes the number of
iterations.

3. Multi-objective optimization
3.1. Pareto optimization
The mix design of concrete should satisfy the concept of sustainable design [68], where the structure should meet its own design
requirements, considering economic efficiency and environmental impact at the same time [69]. Therefore, the mix design of concrete

Table 1
The pseudocode of PSO and GWO.

(a) PSO (b) GWO

for each particle i Initialize the number of wolves Zj (j = 1,2 …,n)

Initialize position xid and velocity vid for particle i Initialize a, A, and C
Evaluate particle i and set Pbesti = Xi Calculate the fitness of each gray wolf
end for Zα = the best gray wolf
Gbest = min {Pbest} Zβ = the second gray wolf
While not stop Zδ = the third gray wolf
for I = 1 to N Zw = the other gray wolf
Updata the position and velocity for particle i while (t < Max number of iterations)
Evaluate particle i for each gray wolf
Updata the fitness of the particle by predict error of PSO-BPNN model Updata the position of the each gray wolf
if fitness (Xi) < fitness (Pbesti) end for
Pbesti = Xi Updata the fitness of the each gray wolf by predict error of GWO-BPNN model
if fitness (Pbesti) < fitness (Gbest) Updata Zα, Zβ, Zδ
Gbest = Pbesti t = t+1
end for end while
end while return Zα as Zbest

8
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

is a typical multi-objective optimization problem. There are usually two ways to solve multi-objective optimization problems: the
priori and posteriori methods [70]. The priori method transforms the multi-objective problem into a single-objective problem by
setting weights in advance, and the disadvantage of this method is that the weights are determined empirically, which will reduce the
objectivity of the optimization results [71]. The posteriori method can solve the problem by determining the Pareto optimal solution
set, use this method to ensure accurate optimization results but it also needs to overcome the complexity of the multi-objective solution
mechanism. Therefore, this section will introduce the basic concepts of multi-objective optimization problems and propose the cor­
responding Pareto optimal solution set. The mathematical model of multi-objective problem is as following:
Min F(x) = {f1 (x), f2 (x), · · ·, fm (x)}
⎧
⎪
⎪ gi (x) ≥ 0, i = 1, 2, · · ·, p
⎨ (15)
s.t. hi (x) = 0, i = 1, 2, · · ·, u
⎪
⎪
⎩
h ≤ xj ≤ l

where F(x) denotes the total function containing objective functions; fm(x) denotes the objective function; gi(x) and hi(x) are the
functions of inequality constraint and the equation constraint, respectively; p and u are the numbers of inequality constraints and
equation constraints, respectively; xj denotes variables of objective function; h and l are the upper and lower bounds of xj, respectively.
Three definitions can be made clear before Pareto optimization including Pareto dominance, Pareto front and Pareto optimal
solution set. Fig. 5 analyzes the above definitions, and the theoretical models of the three definitions are analyzed below.
(a) Pareto Dominate
When a, b1 belong to data aggregation S and supposing {∀i ∈{1, 2, …, n}, fi(a)≤ fi(b1)}∧{∃j∈{1, 2, …, n}, fi(a)≤ fi(b1)} and a
dominates b1, if there is no other decision variable can dominate a, then a is a non-dominated solution.
(b) Pareto front
In the set S of decision variables, for any x belongs to S, if there is a variable x1 such that the objective function F(x1)≤F(x), then x1 is
the absolute optimal solution. If there exists a variable x2 such that the objective function F(x)≤F(x2), then x2 is a non-inferior solution,
and the Pareto front consists of the objective function corresponding to all non-inferior solutions.
(c) Pareto optimal set
a is known as a Pareto optimal solution if there is no decision vector that dominates the decision vector a. The pareto optimal set
consists of all Pareto optimal solutions.

3.2. Technique for Order Preference by similarity to an ideal solution (TOPSIS)

It is easy to find that Pareto optimal set contains multiple Pareto optimal solutions and each solution is equally admissible. In
practical engineering, designers often need to select the optimal RBAC mix design scheme of the optimal solution set. So formulating a
selection rule of Pareto optimal solution is the key to ensure the rationality of mix design. Scholars have proposed evaluation methods
such as Entropy-weight Method (EWM), Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS), analytic hierarchy
process (AHP) etc. In this paper, TOPSIS method is used to determine the optimal mix proportion design scheme for RBAC. The method
is evaluated by calculating the distance between each scheme and the positive ideal scheme (the optimal solution combination of each
objective function) and the negative ideal scheme (the worst solution combination of each objective function):

Fig. 5. The analysis process of pareto optimization.

9
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ n
( )2
di+ = Fi − Fideal (16)
i=1

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ n
( )2
di− = Fi − Fnon− ideal (17)
i=1

di+
score = (18)
di+ + di−

where Fi is the ith solution; Fideal is the positive-ideal solution; Fnon-ideal is the negative-ideal solution; di+ and di- are the distances of a
solution from the positive-ideal solution and the negative-ideal solution, respectively; score is the evaluation score, and the larger the
score is, the closer the solution is to the positive-ideal solution.

4. Database description and analysis

4.1. Variable selection
Although numerous studies have been conducted on mix design of recycled brick aggregate concrete, however, the design pro­
cedure of the currently proposed RBAC mix design method is still complicated. Therefore, this paper collected 182 sets of test data as a
database [40,72–74] and utilized ML model to predict the compressive strength of RBAC. It should be noted that when selecting data,
the authors excluded some parameters such as Plasticizers because these parameters were not applicable to the entire database.
Therefore, the design parameters used in this paper are cement content (Cement), water to cement ratio (W/C), crushed tile ratio (CT),
crushed brick ratio (CB), and natural aggregate (NA) ratio. At the same time, the design parameters selected in this paper are inde­
pendent of each other and the only corresponding concrete mix ratio scheme can be obtained by determining the value of each
parameter. Since the original test does not reflect the source and treatment method of different brick aggregates, the database in this
paper did not analyze the influence of brick aggregate type or treatment method on RBAC compressive strength. Considering the
differences in the effects of coarse and fine aggregates on the compressive strength of RBAC [75], the effects of both on the 28d
compressive strength of RBAC (output variable) are considered separately by dividing fine aggregates (0~4 mm) and coarse aggregates
(4–16 mm). CT, CB and NA satisfy Eq. (19) where the total aggregate volume ratio is 100%:
CT + CB + NA = 100% (19)

4.2. Data characterization

Fig. 6 preliminary analyzes the data distribution characteristics of 8 input variables by establishing scatter matrix. The diagonal line
is the histogram of the frequency distribution of the corresponding variables. The lower left shows the probability density of the
distribution of different variables, and the darker color indicates the greater probability of data appearing in the region. For example,
when W/C takes the value of 0.5, NA (4–16) takes the value of concentrated distribution near 100%. The upper right shows the scatter
plot of different variables to analyze the specific distribution of the data. It can be found that the variables show a trend of nonlinear
distribution, so the linear regression or simple nonlinear regression models cannot accurately analyze the distribution characteristics of
the data and their effects on the output variables. Table .2 analyzes the data characteristics of 9 variables, where the mean values of NA
(0–4) and NA (4–16) are 50% and much higher than that of CT and CB, which corresponds to experimental values of compressive
strength in RBAC studies. In order to fully evaluate the distribution characteristics of input variables and quantify their influence on

Fig. 6. Distribution of data characteristics of input variables.

10
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

Table 2
Description of the database used in this paper.

Parameters No Unit Count Mean Median Min Max SD

3
Cement 1 kg/m 182 368.92 350 180 528 76.60
W/C 2 – 182 0.526 0.5 0.35 1.08 0.11
CT(0–4) 3 % 182 15.86 0 0 100 24.48
CT(4–16) 4 % 182 13.28 0 0 100 22.82
CB(0–4) 5 % 182 18.86 0 0 100 27.13
CB(4–16) 6 % 182 27.85 25 0 100 32.00
NA(0–4) 7 % 182 51.34 50 0 100 36.89
NA(4–16) 8 % 182 53.05 50 0 100 35.57
Compressive strength 1 MPa 182 30.42 25.91 8.7 80.5 14.00

Note: SD denotes standard deviation.

output variables, section 5.3 uses explanatory models (SHAP) to enhance the interpretability of machine learning models and reveals
the mechanism of influence of input variables on output variables.
Before using the ML model, the spearman correlation coefficient is used to analyze the correlation between the variables (see
Fig. 7). It can be found that Cement, W/C and NA (4–16) have a greater correlation with Compressive strength of RBAC, and W/C is
negatively correlated with Compressive strength of RBAC. This phenomenon is also consistent with existing research results [76–78].
For example, the compressive strength of RBAC can be enhanced by appropriately reducing the water-cement ratio and increasing the
content of natural aggregates.

5. Results analysis and discussion

5.1. Parametric analysis of ML model
The performance of ML model is significantly affected by its parameters. It is necessary to optimize the parameters of each ML
model by setting the mean square error (MSE) value in advance. The following methods are used to solve the problem of poor model
prediction accuracy caused by parameter selection:

● The network parameters of BPNN model are adjusted in two ways: (1) Comparing the MSE values generated by using different
number of hidden layers and get the number of hidden layers corresponding to the minimum MSE value as the optimal parameter.
(2) The PSO and GWO algorithms are used to optimize the parameters of the BPNN models. It can be found that the first group of
BPNN models can be used as the control group, and the remaining two groups of BPNN models are used as the experimental group
to compare the optimization ability of the SI algorithm.
● The key parameters of the SVM model are Penalty coefficient c and gamma g, where c controls the distance between the predicted
value and the regression plane. Too large or too small value of c will lead to poor generalization ability of the model. Besides, g
controls the projection ability of the sample to the regression plane. When g takes a larger value, the projection of the sample is
stronger, which can improve the prediction accuracy of the training however. But too strong projection ability can also cause

Fig. 7. Correlation of input and output variables.

11
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

overfitting (good prediction accuracy of the training set and poor prediction accuracy of the testing set) [79]. In this paper, the
libsvm toolbox proposed by Chang et al. [80] is used to perform parameters optimization for c and g. The specific search process can
be found in Fig. 8 (a).
● The key parameters of the RF model are the numbers of leaves and grown trees, and increasing in the number of trees can not only
improve the prediction accuracy but also increase the computational complexity. The number of leaves also determines whether
the model will be over-fitted or not. Two parameters are optimized using pre-set MSE values and the specific process can be seen in
Fig. 8 (b).
● The ELM model is only needed to determine the optimal number of nodes in the hidden layer, so the optimization search of this
model can be done by setting the MSE value.
● The XGB model provided regularization, cross validation and missing value processing, so the model parameters provided by
sklearn in python are used in this paper.
● The GRNN model controls the performance of the model by adjusting the smoothness factor σ of the kernel density function, which
can be used as the optimal parameters of the model by traversing the minimum MSE values to the corresponding values.

5.2. ML predictive performance analysis

Fig. 9 compares and analyzes the predicted and experimental values of RBAC compressive strength obtained by 8 models, where the
predicted and experimental values show a high degree of consistency. Specifically, the ratio of the predicted value to the experimental
value is 0.73–1.16, and the determination coefficient R2 is 0.83–0.99. Among them, the prediction ability of GRNN and ELM models on
the testing set is lower than that of the other 6 models, while GWO-BP model shows good prediction ability in both testing set and
training set. The determination coefficients R2 of the training set and the testing set are 0.97 and 0.99, respectively. The fitting
conditions are y = 0.98x and y = 0.96x, respectively.
The prediction performance of the ML model cannot be effectively evaluated only by individual indexes. Compared with scatter
diagram which can only show the relationship between single indexes, Taylor diagram can evaluate the performance of the ML model
from multiple angles [81]. Therefore, this method is used to evaluate the R, SD and RMSE (Fig. 10) of the Model. It should be noted that
the closer the model points are to the reference point in the Taylor diagram, the better the model performs. For the training set, the
performance of BPNN and ELM models is poor, while the performance of GWO-BP is the best. For the training set, the performance of
ML models can be ordered as GWO-BP > RF > PSO-BP > SVM > XGB > GRNN > ELM > BPNN. For the testing set, the performance of
BPNN is inferior to the other 7 models, while GWO-BP also shows good performance. The performance of ML models can be ordered by
GWO-BP > PSO-BP > XGB > RF > GRNN > SVM > ELM > BPNN. GWO-BP is confirmed to be the most suitable model for predicting
the compressive strength of RBAC by comparing several evaluation indexes of 8 ML models. Meanwhile, comparing BPNN, PSO-BP and
GWO-BP, it can be known that using SI algorithms to optimize the parameters of ML model can effectively improve the prediction
performance and generalization ability of the model.
Fig. 11 analyzes the prediction results of the PSO-BP model, which performed well in predicting the compressive strength of RBAC
on the training and testing sets, with the maximum difference between the predicted value and the experimental value at 21.9Mpa. The
authors thought that the large prediction error between the experimental and predicted values is due to the properties of the recycled
aggregates, and the phenomenon has been similarly concluded in the relevant literature [82]. Other than that, the model predicts well
for more than 90% of the test values.
Compared with PSO, GWO shows better optimization ability (Fig. 12). The maximum error between the predicted value and the
experimental value reduced by 48.9%, GWO-BP could predict well for more than 95% of the test values, and only a few test values were
not correctly predicted. The reason is also that part of the experimental value is beyond the reasonable range.

5.3. Feature importance analysis by SHAP

ML model has limitations due to the existence of black box problem (poor self-interpretation). In this paper, SHapley Additive
exPlanation (SHAP) method is used to solve the problem of poor interpretability of machine learning. This method explains the ML

Fig. 8. Parametric optimization of SVM and RF.

12
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

Fig. 9. Comparison of predicted and experimental values obtained by machine learning.

13
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

model locally and globally by calculating the contribution of each variable to the model [83]. Fig. 13 analyzes the influence of 8 design
parameters on the compressive strength of RBAC. Fig. 13(a) explains the feature importance of the model from a global perspective,
with the vertical axis reflecting the importance of design parameters to the compressive strength of RBAC, the horizontal axis rep­
resents the distribution of SHAP value, and the color bar reflects the variation of the characteristic values of the design parameters.
Therefore, the parameter that had the greatest influence on the compressive strength of RBAC is Cement, and the parameter that had
the least influence is CB(0–4). The specific order is Cement > W/C > NA(4–16)>NA(0–4)>CB(4–16)>CT(4–16)>CT(0–4)>CB(0–4). It
should be noted that the zero value of the horizontal axis is the average value of RBAC compressive strength. For example, for W/C,
SHAP value increases with the decrease of W/C (color changes from red to blue), the compressive strength of RBAC will be 8.07 MPa
lower than its average value when W/C takes the maximum value; and when W/C takes the minimum value, the compressive strength
of RBAC will be 18.76 MPa higher than its average value, since high water-cement ratio means that the concrete has a high porosity
[84], which will significantly reduce the compressive strength of RBAC. Based on hierarchical clustering analysis, Fig. 13(b) analyzes
the influence of design parameters on compressive strength of RBAC from the perspective of a single sample, where horizontal axis
represents the number of samples, vertical axis represents the importance ranking of design parameters, f(x) represents the output
value of the model, the gray line represents the baseline, and the color from blue to red represents the SHAP value from small to large.
For example, Cement, W/C and NA(4–16) can increase the compressive strength of RBAC when the sample number is 175, and CB(0–4)
has less effect on the compressive strength of RBAC.

14
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

Fig. 10. Performance of the ML models on the test and train data.

Fig. 11. Comparison of predicted and experimental values obtained by PSO-BP

Fig. 12. Comparison of predicted and experimental values obtained by GWO-BP.

15
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

Fig. 13. Interpretation of the predicted compressive strength of recycled brick-concrete.

Fig. 14 quantifies the influence of some design parameters on the RBAC compressive strength and the degree of interaction. It can
be found that the SHAP value of Cement increases as the value of Cement increases, while the SHAP value of W/C and CT(4–16)
decreases with the increase of the values of the corresponding parameters. This indicates that Cement is positively correlated with the
compressive strength of RBAC, while W/C and CT(4–16) are negatively correlated with the compressive strength of RBAC. It can also
be found that when the value of the design parameter is constant, the SHAP value of the parameter will be influenced by other design
parameters. For example, when the value of Cement is 400 kg/m3, SHAP value increases with the increase of W/C, but decreases with
the increase of NA (4–16). When the value of CT (4–16) is 25%, SHAP value increases with the increase of NA (4–16). In summary, the
SHAP method can effectively analyze the impact of different design parameters on the compressive strength of RBAC and the degree of
interaction of the variables. Therefore, it can be used to guide the mix design of concrete.

Fig. 14. Local interpretation of variable interactions by SHAP.

16
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

6. Intelligent mix proportion design

6.1. Model building
After obtaining the most suitable ML model for predicting the compressive strength of RBAC, this section builds the MOO model
and develops an optimization model for the RBAC compounding scheme based on multiple perspectives including compressive
strength of RBAC, carbon emission (CO2) and cost (see Fig. 15):
● Determine the design parameters of RBAC mix proportion (input variables of ML model).
● Determine the optimization target as max compressive strength of RBAC, min CO2 and min cost; the constraint conditions are the
upper and lower limits of design parameters.
● Search the Pareto front and the optimal solution set, and use the TOPSIS method to determine the optimal mix proportion of RBAC.
Firstly, the objective function of RBAC mix design model is determined as follows:
⎧
⎨ max compressive strength of RBAC
F(x) = min Cost (20)
⎩
min CO2

Cost = Mc Uc + Mw Uw + MCT(0− 4) UCT(0− 4) + MCT(4− 16) UCT(4− + MCB(0− 4) UCB(0−

(21)
16) 4)
+MCB(4− 16) UCB(4− 16) + MNA(0− 4) UNA(0− 4) + MNA(4− 16) UNA(4− 16)

CO2 = Cc Uc + Cw Uw + CCT(0− 4) UCT(0− 4) + CCT(4− 16) UCT(4− 16) + CCB(0− 4) UCB(0−

(22)
4)
+CCB(4− 16) UCB(4− 16) + CNA(0− 4) UNA(0− 4) + CNA(4− 16) UNA(4− 16)

where Mc, Mw, MCT(0-4), MCT(4-16), MCB(0-4), MCB(4-16), MNA(0-4) and MNA(4-16) are the unit prices of Cement, Water, CT(0–4), CT(4–16),
CB(0–4), CB(4–16), NA(0–4) and NA(4–16), respectively [85]; Cc, Cw, CCT(0-4), CCT(4-16), CCB(0-4), CCB(4-16), CNA(0-4) and CNA(4-16) are
carbon emission factors for Cement, Water, CT(0–4), CT(4–16), CB(0–4), CB(4–16), NA(0–4) and NA(4–16),respectively [86–89] (as
shown in Table .3); Uc, Uw, UCT(0-4), UCT(4-16), UCB(0-4), UCB(4-16), UNA(0-4) and UNA(4-16) are the qualities of Cement, Water, CT(0–4), CT
(4–16), CB(0–4), CB(4–16), NA(0–4) and NA(4–16), respectively. The quality of Water can be obtained from W/C and Cement, while
the qualities of CT, CB and NA can be obtained by converting the ratios into numerical values. The predicted values of compressive
strength of RBAC are obtained by using GWO-BP.
The mix proportion optimization model of RBAC requires setting the value constraints and ratio constraints of design parameters.
The specific constraints are shown in Table .4：

6.2. Results and analysis

This section evaluates the RBAC mix design in the database from local and global perspectives (see Fig. 16). Fig. 16(a)–(d) analyze
CO2-compressive strength, cost-compressive strength, CO2-cost and TOPSIS score for each set of mix proportion, respectively. It can be
found that there is a certain correlation between CO2 and cost corresponding to each group of mix proportion, for the reason that the
highest carbon emission factor and unit price in the RBAC component are both from Cement. Therefore, reducing CO2 and cost of RBAC
matching ratio needs to the control of cement content, which directly affects the compressive strength of RBAC. It seems to be a
conflicting choice, but this is also the key problem that this model expects to solve. The correlation between CO2-compressive strength
and cost-compressive strength is weaker than that of CO2-cost because W/C has a great influence on the compressive strength of RBAC,
while the carbon emission factor and unit price of water are much lower than that of Cement. However, it is not advisable to reduce
costs and carbon emissions by changing the water consumption, because water-cement ratio will affect the performance of RBAC
significantly.
Table .5 gives the top 10 RBAC mix design schemes in the database by using TOPSIS method. It can be found that not all objectives
(max compressive strength of RBAC, min CO2 and min cost) can be optimized at the same time, and each mix proportion shows good
potential on a certain objective. For example, the cost and CO2 of the second group are lower than those of the first group, but the RBAC
compressive strength of this group is weaker than that of the first group. The performance of the target values obtained by the first four
groups of mix proportions is similar, so the corresponding TOPSIS score changes little. It also indicates that under the premise of
ensuring the compressive strength of concrete, it is feasible to appropriately use recycled brick aggregate (25%~75%) to reduce carbon
emissions and costs. Comparing the 9th group with the 10th group, it can be found that the compressive strength and cost of the two
groups are quite different while the corresponding TOPSIS score differs with less extent, which indicates that it is unreasonable to
optimize the RBAC mix proportion by determining the target priority in advance.

7. Conclusion
This paper aims to solve the problems of high cost and complicated procedure of RBAC mix design. Eight machine learning models
are integrated to predict compressive strength of RBAC, and SHAP is used to analyze the influence law of mix design parameters on
compressive strength of RBAC. Finally an intelligent mix design procedure of RBAC based on MOO model is proposed. The following
main conclusions can be drawn:

17
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

Fig. 15. Multi-objective optimization solution of RBAC mix proportion.

Table 3
Cost and carbon emission factors for design parameters.

Parameters Symbols Cost ($/kg) Carbon emission factors (kg CO2 e)

Cement content (kg/m3) Cement 0.084 0.735

Water (kg) W 0.00057 0.000168
Crushed tile (0–4) (kg) CT(0–4) 0.0063 0.00126
Crushed tile (4–16) (kg) CT(4–16) 0.0077 0.0026
Crushed brick (0–4) (kg) CB(0–4) 0.0063 0.00126
Crushed brick (4–16) (kg) CB(4–16) 0.0077 0.0026
Natural aggregate (0–4) (kg) NA(0–4) 0.021 0.00251
Natural aggregate (4–16) (kg) NA(4–16) 0.017 0.00523

Table 4
Constraints of the mix proportion model in this paper.

Parameters Symbols Upper Lower

Cement content (kg) Cement 528 180

Water-cement ratio (− ) W/C 1.08 0.35
Crushed tile ratio (0–4) (%) CT(0–4) 100 0
Crushed tile ratio (4–16) (%) CT(4–16) 100 0
Crushed brick ratio (0–4) (%) CB(0–4) 100 0
Crushed brick ratio (4–16) (%) CB(4–16) 100 0
Natural aggregate ratio (0–4) (%) NA(0–4) 100 0
Natural aggregate ratio (4–16) (%) NA(4–16) 100 0
Compressive strength of RBAC (MPa) Compressive strength 80.5 8.7

(1) ML model based on SI optimization has better prediction performance than traditional ML model. The predicted/tested values
using PSO-BP and GWO-BP in the training set (testing set) are 0.91 (0.90) and 0.98 (0.96), respectively. The most suitable model
for RBAC compressive strength prediction is GWO-BP.
(2) Cement and W/C are the key parameters affecting the compressive strength of RBAC. When the upper and lower limits of
Cement and W/C are taken respectively, the compressive strength of RBAC is reduced by 29.40Mpa and 26.83 MPa, respec­
tively. The impact of coarse aggregate on the compressive strength of RBAC is about 1.47 times that of fine aggregate, which is
consistent with the experimental results.
(3) The multi-objective optimization model considering RBAC compressive strength, carbon emission and cost can effectively guide
RBAC mix design. There is a certain correlation between CO2, compressive strength and cost. Cement is the key parameter
affecting RBAC carbon emissions and costs. Under the premise of ensuring the compressive strength of concrete, using recycled
brick aggregate to replace natural aggregate can effectively reduce carbon emissions and costs, and the replacement rate of
recycled brick aggregate ranges from 25%~75%.

18
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

Fig. 16. Pareto front showing the trade-off among Cost, compressive strength and CO2 for recycled brick-concrete.

Table 5
Constraints of the mix proportion model in this paper.

No. Cement W/C CT CT CB CB NA NA Compressive Cost CO2 TOPSIS

(0–4) (4–16) (0–4) (4–16) (0–4) (4–16) strength score

1 180 0.35 0 0 0 25 0 75 64.53 31.57 137.39 100

2 180 0.43 0 0 0 75 0 25 58.59 26.48 135.94 99.30
3 180 0.35 0 0 0 50 0 50 60.01 29.02 136.66 98.96
4 180 0.4 0 0 0 50 0 50 59.10 29.03 136.70 98.23
5 180 0.35 0 0 0 0 0 100 59.58 34.11 138.11 95.35
6 180 0.46 50 0 0 50 0 0 49.74 22.02 134.23 93.07
7 180 0.377 0 0 0 25 0 75 50.27 31.57 137.38 89.69
8 340 0.45 75 0 0 25 0 0 23.77 26.18 177.85 71.26
9 340 0.4 0 0 0 0 30 70 31.14 38.37 181.04 69.27
10 340 0.45 50 0 0 50 0 0 20.09 27.13 178.33 68.79

(4) In future work, the database should be expanded and the influence of design parameters on the working performance, cost and
carbon emissions of concrete will be further studied to determine the correlation and priority between multiple objectives to
provide more effective mix design solutions.

Credit author statement

Shiqi Wang: Conceptualization, Software, Writing - Original Draft; Peng Xia: Conceptualization, Writing - Original Draft; Zhao
Wang: Formal analysis; Tao Meng: Data Curation; Fuyuan Gong: Writing - Review & Editing, Supervision.

Declaration of competing interest

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

Data availability

Data will be made available on request.

Acknowledgements
This work was sponsored by the “Pioneer” and “Leading Goose” R&D Program of Zhejiang (2023C04042) and National Natural
Science Foundation of China (52008367).

19
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

References
[1] A.A. Aliabdo, A.-E.M. Abd-Elmoaty, H.H. Hassan, Utilization of crushed clay brick in concrete industry, Alex. Eng. J. 53 (1) (2014) 151–168, https://doi.org/
10.1016/j.aej.2013.12.003.
[2] X. He, Z. Zheng, J. Yang, Y. Su, T. Wang, B. Strnadel, Feasibility of incorporating autoclaved aerated concrete waste for cement replacement in sustainable
building materials, J. Clean. Prod. 250 (2020), 119455, https://doi.org/10.1016/j.jclepro.2019.119455.
[3] Q. Tang, Z. Ma, H. Wu, W. Wang, The utilization of eco-friendly recycled powder from concrete and brick waste in new concrete: a critical review, Cem. Concr.
Compos. 114 (2020), 103807, https://doi.org/10.1016/j.cemconcomp.2020.103807.
[4] W. Hawkins, J. Orr, T. Ibell, P. Shepherd, A design methodology to reduce the embodied carbon of concrete buildings using thin-shell floors, Eng. Struct. 207
(2020), 110195, https://doi.org/10.1016/j.engstruct.2020.110195.
[5] L. Evangelista, J. de Brito, Mechanical behaviour of concrete made with fine recycled concrete aggregates, Cem. Concr. Compos. 29 (5) (2007) 397–401, https://
doi.org/10.1016/j.cemconcomp.2006.12.004.
[6] K. Rahal, Mechanical properties of concrete with recycled coarse aggregate, Build. Environ. 42 (1) (2007) 407–415, https://doi.org/10.1016/j.
buildenv.2005.07.033.
[7] W. Lu, H. Yuan, A framework for understanding waste management studies in construction, Waste Manage. (Tucson, Ariz.) 31 (6) (2011) 1252–1260, https://
doi.org/10.1016/j.wasman.2011.01.018.
[8] A. Barbudo, J. de Brito, L. Evangelista, M. Bravo, F. Agrela, Influence of water-reducing admixtures on the mechanical performance of recycled concrete,
J. Clean. Prod. 59 (2013) 93–98, https://doi.org/10.1016/j.jclepro.2013.06.022.
[9] S. Silva, L. Evangelista, J. de Brito, Durability and shrinkage performance of concrete made with coarse multi-recycled concrete aggregates, Construct. Build.
Mater. 272 (2021), 121645, https://doi.org/10.1016/j.conbuildmat.2020.121645.
[10] L. Zhu, J. Dai, G. Bai, F. Zhang, Study on thermal properties of recycled aggregate concrete and recycled concrete blocks, Construct. Build. Mater. 94 (2015)
620–628, https://doi.org/10.1016/j.conbuildmat.2015.07.058.
[11] K.K. Sagoe-Crentsil, T. Brown, A.H. Taylor, Performance of concrete made with commercially produced coarse recycled concrete aggregate, Cement Concr. Res.
31 (5) (2001) 707–712, https://doi.org/10.1016/S0008-8846(00)00476-2.
[12] S.C. Kou, C.S. Poon, M. Etxeberria, Influence of recycled aggregates on long term mechanical properties and pore size distribution of concrete, Cem. Concr.
Compos. 33 (2) (2011) 286–291, https://doi.org/10.1016/j.cemconcomp.2010.10.003.
[13] C. Wu, Z. Hong, J. Zhang, S.C. Kou, Pore size distribution and ITZ performance of mortars prepared with different bio-deposition approaches for the treatment of
recycled concrete aggregate, Cem. Concr. Compos. 111 (2020), 103631, https://doi.org/10.1016/j.cemconcomp.2020.103631.
[14] C. Thomas, J. Setién, J.A. Polanco, A.I. Cimentada, C. Medina, Influence of curing conditions on recycled aggregate concrete, Construct. Build. Mater. 172
(2018) 618–625, https://doi.org/10.1016/j.conbuildmat.2018.04.009.
[15] P. Pereira, L. Evangelista, J. de Brito, The effect of superplasticisers on the workability and compressive strength of concrete made with fine recycled concrete
aggregates, Construct. Build. Mater. 28 (1) (2012) 722–729, https://doi.org/10.1016/j.conbuildmat.2011.10.050.
[16] C. Zhou, Z. Chen, Mechanical properties of recycled concrete made with different types of coarse aggregate, Construct. Build. Mater. 134 (2017) 497–506,
https://doi.org/10.1016/j.conbuildmat.2016.12.163.
[17] M. Rezaiee Pajand, J. Mohebi Najm Abad, A. Karimipour, A. Rezaiee-Pajand, Propose new implement models to determine the compressive, tensile and flexural
strengths of recycled coarse aggregate concrete via imperialist competitive algorithm, J. Build. Eng. 40 (2021), 102337, https://doi.org/10.1016/j.
jobe.2021.102337.
[18] J. Xiao, Y. Fan, M.M. Tawana, Residual compressive and flexural strength of a recycled aggregate concrete following elevated temperatures, Struct. Concr. 14
(2) (2013) 168–175, https://doi.org/10.1002/suco.201200037.
[19] H.J. Ho, A. Iizuka, E. Shibata, Chemical recycling and use of various types of concrete waste: a review, J. Clean. Prod. 284 (2021), 124785, https://doi.org/
10.1016/j.jclepro.2020.124785.
[20] B. Bhardwaj, P. Kumar, Waste foundry sand in concrete: a review, Construct. Build. Mater. 156 (2017) 661–674, https://doi.org/10.1016/j.
conbuildmat.2017.09.010.
[21] Q. Wang, Y. Geng, Y. Wang, H. Zhang, Drying shrinkage model for recycled aggregate concrete accounting for the influence of parent concrete, Eng. Struct. 202
(2020), 109888, https://doi.org/10.1016/j.engstruct.2019.109888.
[22] J.M. Khatib, Properties of concrete incorporating fine recycled aggregate, Cement Concr. Res. 35 (4) (2005) 763–769, https://doi.org/10.1016/j.
cemconres.2004.06.017.
[23] Y. Wang, H. Zhang, Y. Geng, Q. Wang, S. Zhang, Prediction of the elastic modulus and the splitting tensile strength of concrete incorporating both fine and
coarse recycled aggregate, Construct. Build. Mater. 215 (2019) 332–346, https://doi.org/10.1016/j.conbuildmat.2019.04.212.
[24] M. Pepe, R.D. Toledo Filho, E.A.B. Koenders, E. Martinelli, A novel mix design methodology for Recycled Aggregate Concrete, Construct. Build. Mater. 122
(2016) 362–372, https://doi.org/10.1016/j.conbuildmat.2016.06.061.
[25] S. Pradhan, S. Kumar, S.V. Barai, Recycled aggregate concrete: particle Packing Method (PPM) of mix design approach, Construct. Build. Mater. 152 (2017)
269–284, https://doi.org/10.1016/j.conbuildmat.2017.06.171.
[26] M.S. Bidabadi, M. Akbari, O. Panahi, Optimum mix design of recycled concrete based on the fresh and hardened properties of concrete, J. Build. Eng. 32 (2020),
101483, https://doi.org/10.1016/j.jobe.2020.101483.
[27] X. Ding, J. Hao, Z. Chen, J. Qi, M. Marco, New mix design method for recycled concrete using mixed source concrete coarse aggregate, Waste and Biomass
Valorization 11 (10) (2020) 5431–5443, https://doi.org/10.1007/s12649-020-01073-7.
[28] X. Liu, J. Wu, P. Yan, X. Zhao, A novel mix design method for mixed recycled coarse aggregate concrete, J. Mater. Civ. Eng. 34 (5) (2022), 04022071, https://
doi.org/10.1061/(ASCE)MT.1943-5533.0004213.
[29] D. Meng, X. Wu, H. Quan, C. Zhu, A strength-based mix design method for recycled aggregate concrete and consequent durability performance, Construct. Build.
Mater. 281 (2021), 122616, https://doi.org/10.1016/j.conbuildmat.2021.122616.
[30] A. Hammoudi, K. Moussaceb, C. Belebchouche, F. Dahmoune, Comparison of artificial neural network (ANN) and response surface methodology (RSM)
prediction in compressive strength of recycled concrete aggregates, Construct. Build. Mater. 209 (2019) 425–436, https://doi.org/10.1016/j.
conbuildmat.2019.03.119.
[31] M.R. Kaloop, A.R. Gabr, S.M. El Badawy, A. Arisha, S. Shwally, J. Hu, Predicting resilient modulus of recycled concrete and clay masonry blends for pavement
applications using soft computing techniques, Front. Struct. Civ. Eng. 13 (6) (2019) 1379–1392, https://doi.org/10.1007/s11709-019-0562-2.
[32] E.M. Golafshani, A. Behnood, Application of soft computing methods for predicting the elastic modulus of recycled aggregate concrete, J. Clean. Prod. 176
(2018) 1163–1176, https://doi.org/10.1016/j.jclepro.2017.11.186.
[33] S. Arora, B. Singh, B. Bhardwaj, Strength performance of recycled aggregate concretes containing mineral admixtures and their performance prediction through
various modeling techniques, J. Build. Eng. 24 (2019), 100741, https://doi.org/10.1016/j.jobe.2019.100741.
[34] M. Amiri, F. Hatami, Prediction of mechanical and durability characteristics of concrete including slag and recycled aggregate concrete with artificial neural
networks (ANNs), Construct. Build. Mater. 325 (2022), 126839, https://doi.org/10.1016/j.conbuildmat.2022.126839.
[35] K. Liu, C. Zou, X. Zhang, J. Yan, Innovative prediction models for the frost durability of recycled aggregate concrete using soft computing methods, J. Build. Eng.
34 (2021), 101822, https://doi.org/10.1016/j.jobe.2020.101822.
[36] M. Cheng, D. Prayogo, Y. Wu, Novel genetic algorithm-based evolutionary support vector machine for optimizing high-performance concrete mixture,
J. Comput. Civ. Eng. 28 (4) (2014), 06014003, https://doi.org/10.1061/(ASCE)CP.1943-5487.0000347.
[37] J. Zhang, Y. Huang, Y. Wang, G. Ma, Multi-objective optimization of concrete mixture proportions using machine learning and metaheuristic algorithms,
Construct. Build. Mater. 253 (2020), 119208, https://doi.org/10.1016/j.conbuildmat.2020.119208.

20
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

[38] J. Zhang, Y. Huang, G. Ma, Y. Yuan, B. Nener, Automating the mixture design of lightweight foamed concrete using multi-objective firefly algorithm and support
vector regression, Cem. Concr. Compos. 121 (2021), 104103, https://doi.org/10.1016/j.cemconcomp.2021.104103.
[39] J. Sun, X. Wang, J. Zhang, F. Xiao, Y. Sun, Z. Ren, G. Zhang, S. Liu, Y. Wang, Multi-objective optimisation of a graphite-slag conductive composite applying a
BAS-SVR based model, J. Build. Eng. 44 (2021), 103223, https://doi.org/10.1016/j.jobe.2021.103223.
[40] T. Kalman Šipoš, I. Miličević, R. Siddique, Model for mix design of brick aggregate concrete based on neural network modelling, Construct. Build. Mater. 148
(2017) 757–769, https://doi.org/10.1016/j.conbuildmat.2017.05.111.
[41] R. Kumar Tipu, V.R. Panchal, K.S. Pandya, An ensemble approach to improve BPNN model precision for predicting compressive strength of high-performance
concrete, Structures 45 (2022) 500–508, https://doi.org/10.1016/j.istruc.2022.09.046.
[42] C. Jiang, G. Liang, Modeling shear strength of medium- to ultra-high-strength concrete beams with stirrups using SVR and genetic algorithm, Soft Comput. 25
(16) (2021) 10661–10675, https://doi.org/10.1007/s00500-021-06027-2.
[43] Z.M. Yaseen, M.T. Tran, S. Kim, T. Bakhshpoori, R.C. Deo, Shear strength prediction of steel fiber reinforced concrete beam using hybrid intelligence models: a
new approach, Eng. Struct. 177 (2018) 244–255, https://doi.org/10.1016/j.engstruct.2018.09.074.
[44] B. Keshtegar, M. Bagheri, Z.M. Yaseen, Shear strength of steel fiber-unconfined reinforced concrete beam simulation: application of novel intelligent model,
Compos. Struct. 212 (2019) 230–242, https://doi.org/10.1016/j.compstruct.2019.01.004.
[45] M. Muthukumar, D. Mohan, M. Rajendran, Optimization of mix proportions of mineral aggregates using Box Behnken design of experiments, Cem. Concr.
Compos. 25 (7) (2003) 751–758, https://doi.org/10.1016/S0958-9465(02)00116-6.
[46] M.S. Ahmad, S.M. Adnan, S. Zaidi, P. Bhargava, A novel support vector regression (SVR) model for the prediction of splice strength of the unconfined beam
specimens, Construct. Build. Mater. 248 (2020), 118475, https://doi.org/10.1016/j.conbuildmat.2020.118475.
[47] H. Chen, T. Deng, T. Du, B. Chen, M.J. Skibniewski, L. Zhang, An RF and LSSVM–NSGA-II method for the multi-objective optimization of high-performance
concrete durability, Cem. Concr. Compos. 129 (2022), 104446, https://doi.org/10.1016/j.cemconcomp.2022.104446.
[48] S.M.S.M.K. Samarakoon, P. Ruben, J. Wie Pedersen, L. Evangelista, Mechanical performance of concrete made of steel fibers from tire waste, Case Stud. Constr.
Mater. 11 (2019), e00259, https://doi.org/10.1016/j.cscm.2019.e00259.
[49] A.K. Al Shamiri, J. Kim, T. Yuan, Y. Yoon, Modeling the compressive strength of high-strength concrete: an extreme learning approach, Construct. Build. Mater.
208 (2019) 204–219, https://doi.org/10.1016/j.conbuildmat.2019.02.165.
[50] M. Li, J. Wang, An empirical comparison of multiple linear regression and artificial neural network for concrete dam deformation modelling, Math. Probl Eng.
(2019), 7620948, https://doi.org/10.1155/2019/7620948, 2019.
[51] A. Dhini, I. Surjandari, B. Kusumoputro, A. Kusiak, Extreme learning machine – radial basis function (ELM-RBF) networks for diagnosing faults in a steam
turbine, J. Ind. Prod. Eng. 39 (7) (2022) 572–580, https://doi.org/10.1080/21681015.2021.1887948.
[52] N.V. Luat, S.W. Han, K. Lee, Genetic algorithm hybridized with eXtreme gradient boosting to predict axial compressive capacity of CCFST columns, Compos.
Struct. 278 (2021), 114733, https://doi.org/10.1016/j.compstruct.2021.114733.
[53] M. Dissanayake, H. Nguyen, K. Poologanathan, G. Perampalam, I. Upasiri, H. Rajanayagam, T. Suntharalingam, Prediction of shear capacity of steel channel
sections using machine learning algorithms, Thin-Walled Struct. 175 (2022), 109152, https://doi.org/10.1016/j.tws.2022.109152.
[54] L. Qiao, Z. Wang, J. Zhu, Application of improved GRNN model to predict interlamellar spacing and mechanical properties of hypereutectoid steel, Mater. Sci.
Eng. 792 (2020), 139845, https://doi.org/10.1016/j.msea.2020.139845.
[55] U. Kurup Pradeep, P. Griffin Erin, Prediction of soil composition from CPT data using general regression neural network, J. Comput. Civ. Eng. 20 (4) (2006)
281–289, https://doi.org/10.1061/(ASCE)0887-3801(2006)20:4(281).
[56] J. J, P. Samui, D. Kim, Reliability analysis of simply supported beam using GRNN, ELM and GPR, Struct. Eng. Mech. 71 (2019) 739–749, https://doi.org/
10.12989/sem.2019.71.6.739.
[57] S. Kim, V.P. Singh, C. Lee, Y. Seo, Modeling the physical dynamics of daily dew point temperature using soft computing techniques, KSCE J. Civ. Eng. 19 (6)
(2015) 1930–1940, https://doi.org/10.1007/s12205-014-1197-4.
[58] F. Sadeghi, Y. Yu, X. Zhu, J. Li, Damage identification of steel-concrete composite beams based on modal strain energy changes through general regression
neural network, Eng. Struct. 244 (2021), 112824, https://doi.org/10.1016/j.engstruct.2021.112824.
[59] S. Yelisetti, V.K. Saini, R. Kumar, R. Lamba, A. Saxena, Optimal energy management system for residential buildings considering the time of use price with
swarm intelligence algorithms, J. Build. Eng. 59 (2022), 105062, https://doi.org/10.1016/j.jobe.2022.105062.
[60] J. Tang, G. Liu, Q. Pan, A review on representative swarm intelligence algorithms for solving optimization problems: applications and trends, IEEE/CAA Journal
of Automatica Sinica 8 (10) (2021) 1627–1643, https://doi.org/10.1109/JAS.2021.1004129.
[61] B.A. Salami, M. Iqbal, A. Abdulraheem, F.E. Jalal, W. Alimi, A. Jamal, T. Tafsirojjaman, Y. Liu, A. Bardhan, Estimating compressive strength of lightweight
foamed concrete using neural, genetic and ensemble machine learning approaches, Cem. Concr. Compos. 133 (2022), 104721, https://doi.org/10.1016/j.
cemconcomp.2022.104721.
[62] B. Chen, C. Zhou, Y. Liu, J. Liu, Correlation analysis of runway icing parameters and improved PSO-LSSVM icing prediction, Cold Reg. Sci. Technol. 193 (2022),
103415, https://doi.org/10.1016/j.coldregions.2021.103415.
[63] S. Khatir, S. Tiachacht, C.-L. Thanh, T.Q. Bui, M. Abdel Wahab, Damage assessment in composite laminates using ANN-PSO-IGA and Cornwell indicator,
Compos. Struct. 230 (2019), 111509, https://doi.org/10.1016/j.compstruct.2019.111509.
[64] R. Poli, J. Kennedy, T. Blackwell, Particle swarm optimization, Swarm. Intell 1 (1) (2007) 33–57, https://doi.org/10.1007/s11721-007-0002-0.
[65] W. Liu, A. Li, C. Liu, Multi-objective optimization control for tunnel boring machine performance improvement under uncertainty, Autom. ConStruct. 139
(2022), 104310, https://doi.org/10.1016/j.autcon.2022.104310.
[66] H. Yang, Z. Wang, K. Song, A new hybrid grey wolf optimizer-feature weighted-multiple kernel-support vector regression technique to predict TBM
performance, Eng. Comput. 38 (3) (2022) 2469–2485, https://doi.org/10.1007/s00366-020-01217-2.
[67] N.T. Ngo, H.A. Le, T.P.T. Pham, Integration of support vector regression and grey wolf optimization for estimating the ultimate bearing capacity in concrete-
filled steel tube columns, Neural Comput. Appl. 33 (14) (2021) 8525–8542, https://doi.org/10.1007/s00521-020-05605-z.
[68] B. Xia, J. Xiao, T. Ding, K. Zhang, Probabilistic sustainability design of structural concrete components under climate change, Struct. Saf. 92 (2021), 102103,
https://doi.org/10.1016/j.strusafe.2021.102103.
[69] B. Xia, J. Xiao, S. Li, Sustainability-based reliability design for reuse of concrete components, Struct. Saf. 98 (2022), 102241, https://doi.org/10.1016/j.
strusafe.2022.102241.
[70] J. Branke, T. Kaußler, H. Schmeck, Guidance in evolutionary multi-objective optimization, Adv. Eng. Software 32 (6) (2001) 499–507, https://doi.org/10.1016/
S0965-9978(00)00110-1.
[71] A. Messac, C.A. Mattson, Generating well-distributed sets of pareto points for engineering design using physical programming, Optim. Eng. 3 (4) (2002)
431–450, https://doi.org/10.1023/A:1021179727569.
[72] C. Zheng, C. Lou, G. Du, X. Li, Z. Liu, L. Li, Mechanical properties of recycled concrete with demolished waste concrete aggregate and clay brick aggregate,
Results. Phys 9 (2018) 1317–1322, https://doi.org/10.1016/j.rinp.2018.04.061.
[73] J. Dang, J. Zhao, Influence of waste clay bricks as fine aggregate on the mechanical and microstructural properties of concrete, Construct. Build. Mater. 228
(2019), 116757, https://doi.org/10.1016/j.conbuildmat.2019.116757.
[74] S.C. Paul, A.J. Babafemi, V. Anggraini, M.M. Rahman, Properties of Normal and Recycled Brick Aggregates for Production of Medium Range (25–30 MPa)
Structural Strength Concrete, Buildings, 2018.
[75] U. Mohammed Tarek, A. Hasnat, A. Awal Mohammad, Z. Bosunia Shamim, Recycling of brick aggregate concrete as coarse aggregate, J. Mater. Civ. Eng. 27 (7)
(2015) B4014005, https://doi.org/10.1061/(ASCE)MT.1943-5533.0001043.
[76] C. Shi, K. Zheng, A review on the use of waste glasses in the production of cement and concrete, Resour. Conserv. Recycl. 52 (2) (2007) 234–247, https://doi.
org/10.1016/j.resconrec.2007.01.013.

21
S. Wang et al. Journal of Building Engineering 71 (2023) 106508

[77] M. Sahmaran, O. Kasap, K. Duru, I.O. Yaman, Effects of mix composition and water–cement ratio on the sulfate resistance of blended cements, Cem. Concr.
Compos. 29 (3) (2007) 159–167, https://doi.org/10.1016/j.cemconcomp.2006.11.007.
[78] Z. Lafhaj, M. Goueygou, A. Djerbi, M. Kaczmarek, Correlation between porosity, permeability and ultrasonic parameters of mortar with variable water/cement
ratio and water content, Cement Concr. Res. 36 (4) (2006) 625–633, https://doi.org/10.1016/j.cemconres.2005.11.009.
[79] O. Devos, C. Ruckebusch, A. Durand, L. Duponchel, J.-P. Huvenne, Support vector machines (SVM) in near infrared (NIR) spectroscopy: focus on parameters
optimization and model interpretation, Chemometr. Intell. Lab. Syst. 96 (1) (2009) 27–33, https://doi.org/10.1016/j.chemolab.2008.11.005.
[80] C.C. Chang, C.J. Lin, Training nu-support vector classifiers: theory and algorithms, Neural Comput. 13 (9) (2001) 2119–2147, https://doi.org/10.1162/
089976601750399335.
[81] C. Bai, H. Nguyen, P.G. Asteris, T. Nguyen-Thoi, J. Zhou, A refreshing view of soft computing models for predicting the deflection of reinforced concrete beams,
Appl. Soft Comput. 97 (2020), 106831, https://doi.org/10.1016/j.asoc.2020.106831.
[82] A. Khademi, K. Behfarnia, T. Kalman Šipoš, I. Miličević, The use of machine learning models in estimating the compressive strength of recycled brick aggregate
concrete, Computational Engineering and Physical Modeling 4 (4) (2021) 1–25, https://doi.org/10.22115/cepm.2021.297016.1181.
[83] B. Hilloulin, V.Q. Tran, Using machine learning techniques for predicting autogenous shrinkage of concrete incorporating superabsorbent polymers and
supplementary cementitious materials, J. Build. Eng. 49 (2022), 104086, https://doi.org/10.1016/j.jobe.2022.104086.
[84] H.S. Wong, N.R. Buenfeld, Determining the water–cement ratio, cement content, water content and degree of hydration of hardened cement paste: method
development and validation on paste samples, Cement Concr. Res. 39 (10) (2009) 957–965, https://doi.org/10.1016/j.cemconres.2009.06.013.
[85] J. Liu, Y. Liu, X. Wang, An environmental assessment model of construction and demolition waste based on system dynamics: a case study in Guangzhou,
Environ. Sci. Pollut. Res. 27 (30) (2020) 37237–37259, https://doi.org/10.1007/s11356-019-07107-5.
[86] J. Xiao, C. Wang, T. Ding, A. Akbarnezhad, A recycled aggregate concrete high-rise building: structural performance and embodied carbon footprint, J. Clean.
Prod. 199 (2018) 868–881, https://doi.org/10.1016/j.jclepro.2018.07.210.
[87] C. Chen, G. Habert, Y. Bouzidi, A. Jullien, A. Ventura, LCA allocation procedure used as an incitative method for waste recycling: an application to mineral
additions in concrete, Resour. Conserv. Recycl. 54 (12) (2010) 1231–1240, https://doi.org/10.1016/j.resconrec.2010.04.001.
[88] P. Hájek, C. Fiala, M. Kynčlová, Life cycle assessments of concrete structures – a step towards environmental savings, Struct. Concr. 12 (1) (2011) 13–22,
https://doi.org/10.1002/suco.201000026.
[89] J. Xiao, Y. Xiao, Y. Liu, T. Ding, Carbon emission analyses of concretes made with recycled materials considering CO2 uptake through carbonation absorption,
Struct. Concr. 22 (S1) (2021) E58–E73, https://doi.org/10.1002/suco.201900577.

22