Journal of Environmental Management 371 (2024) 123068

Contents lists available at ScienceDirect

Journal of Environmental Management

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

Research article

Improved prediction accuracy for compressive strength of recycled

aggregate concrete using optimization-based algorithms and cascade
forward neural network
Xiaoguang Zhou a,b , Jian Zhou a,b,c , Joy P. Ohl d,*
a
State Key Laboratory of Precision Blasting, Jianghan University, Wuhan, 430056, China
b
Hubei Key Laboratory of Blasting Engineering, Jianghan University, Wuhan, 430056, China
c
School of Resources and Safety Engineering, Central South University, Changsha, Hunan, 410083, China
d
NanoStructuraCivis Center: Advancing Smart Materials in Civil Engineering, Tokyo, Japan

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

Handling editor: Lixiao Zhang This study proposed a data driven approach to predict the compressive strength (CS) of recycled aggregate
concrete (RAC) for sustainable construction using an elite single genetic optimization algorithm-based cascade
Keywords: forward neural network (ESGA-CFNN) model. It was applied to 272 RAC samples under different conditions and
Sustainable building materials compositions focusing on key parameters for CS prediction: water-to-cement ratio (WCR), water absorption
Compressive strength prediction
(WA), recycled coarse aggregate (RCA) density, fine aggregate (FA) density, naturally occurring coarse aggregate
Recycled aggregate concrete (RAC)
(NCA) density and water-to-total material ratio (WTMR). These parameters were used to develop the ESGA-
Machine learning
Elite single genetic optimization algorithm CFNN model which was then evaluated for its performance. To compare the ESGA-CFNN model, two other
Cascade forward neural network models were developed and compared: particle swarm optimization-based CFNN (PSO-CFNN) and artificial bee
colony-based CFNN (ABC-CFNN). K-fold cross-validation was used during model development to prevent over-
fitting. Results showed that ESGA-CFNN model performed better with an RMSE (root-mean-squared error) of
1.144, R2 (determination coefficient) of 0.991 and a10-index of 1.000. ABC-CFNN model had an RMSE of 1.434,
R2 of 0.987 and a10-index of 0.982 while PSO-CFNN had an RMSE of 1.561, R2 of 0.984 and a10-index of 0.982.
Practical validation with 6 RAC samples confirmed the real world applicability of these models. The findings of
this study showed that the proposed ESGA-CFNN model is important for quality control in RAC production and
optimizing mix designs to achieve required compressive strength to meet standards and reduce cost and increase
sustainability in concrete construction. This study introduces a novel hybrid approach combining ESGA-CFNN,
PSO-CFNN, and ABC-CFNN algorithms for accurately predicting the compressive strength of RAC. These
models outperform traditional methodologies by offering enhanced predictive accuracy and generalization
capability, especially in complex, real-world datasets.

1. Introduction to purchase new materials and lowers disposal costs (Tam et al., 2018).
Moreover, properly processed RAC can achieve strength comparable to
Recycled aggregate concrete (RAC) is a sustainable construction that of conventional concrete (Nedeljković et al., 2021; de Andrade
material that addresses environmental concerns by reducing the demand Salgado and de Andrade Silva, 2022). With increasing regulatory pres-
for natural resources and minimizing construction waste (Akhtar and sure and sustainability goals, the use of recycled materials in construc-
Sarmah, 2018). RAC is produced from recycled materials obtained from tion is being promoted globally (Su et al., 2020).
demolished concrete structures, which are processed to remove con- However, achieving consistent compressive strength (CS) in RAC
taminants and then crushed into aggregates. This process helps conserve remains challenging due to the variability in recycled aggregate prop-
natural resources and reduces the burden on landfills (Kabirifar et al., erties, which can affect the overall performance of the concrete (Serres
2020; Tang et al., 2020). et al., 2016). This variability highlights the importance of developing
The advantages of RAC include cost savings, as it decreases the need accurate models to predict the CS of RAC, which are essential for

* Corresponding author. NanoStructuraCivis Lab: Advancing Smart Materials in Civil Engineering, Tokyo, Japan.
E-mail address: [email protected] (J.P. Ohl).

https://doi.org/10.1016/j.jenvman.2024.123068
Received 22 August 2024; Received in revised form 16 October 2024; Accepted 21 October 2024
0301-4797/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 1. Process of obtaining RCA.

optimizing mix designs and ensuring the safe and efficient use of RAC in Revilla-Cuesta et al. (2022b) demonstrated the potential of these
construction projects (Wang et al., 2012; Zhu et al., 2018; Piyaphipat methods for estimating compressive strength and modulus of elasticity
et al., 2023). in self-compacting concrete (SCC) containing recycled aggregates, while
Alwash et al. (2017) and Boussahoua et al. (2023) highlighted the need
2. Literature review for improved accuracy in combining NDT with destructive core testing
for concrete strength assessment.
Recent advancements in machine learning (ML) and metaheuristic The use of regression models with density correction has been shown
algorithms have shown promise in enhancing predictions and optimi- to yield precise predictions of elastic modulus in concrete, as seen in the
zations in geotechnical engineering (Zhou et al., 2022b,a, 2023), ma- work by Revilla-Cuesta et al. (2022a), further supporting the integration
terials and civil engineering (Liao and Li, 2020; Nguyen et al., 2020, of regression-based ML models in predicting mechanical properties of
2024; Song et al., 2022; Yonar and Yonar, 2023), and they also have concrete composites.
significantly improved the prediction of concrete composite properties, Regarding the use of artificial intelligence (AI) techniques to predict
offering robust alternatives to traditional empirical methods. P. Zhang the CS of RAC, Z. H. Duan et al. (2013) utilized an artificial neural
et al. (2024) reviewed various ML models, such as artificial neural network (ANN) model with 14 input variables and 146 samples to pre-
networks (ANN) and fuzzy inference systems, for predicting geopolymer dict the CS of RAC. Deshpande et al. (2014) utilized the ANN model, tree
composite properties, emphasizing the importance of large datasets for model, and nonlinear regression analysis to forecast the CS of RAC, with
accurate predictions. Similarly, X. Zhang et al. (2024) applied particle the ANN model proving to be the most efficient. Abdollahzadeh et al.
swarm optimization-enhanced backpropagation neural networks (2016) also proposed 20 models using gene expression programming
(PSO-BPNN) to predict the durability of geopolymer mortar, demon- (GEP) for predicting CS of RAC based on experimental data of 228
strating improved accuracy in forecasting durability-related properties specimens with promising results. Deng et al. (2018) utilized deep
compared to traditional methods. learning theory to predict the CS of RAC based on concrete block ma-
Furthermore, the application of ML techniques like least squares sonry with different mix ratios, and their study showed higher precision
support vector regression (LS-SVR) has shown promise in predicting than the ANN model for the same purpose. Hammoudi et al. (2019)
rheological parameters and flowability in alkali-activated composites, as compared the ANN and response surface model (RSM) for modeling CS
demonstrated by Zhang et al. (2022). Liu et al. (2021) extended this of RAC at different ages and concluded that the ANN model was more
approach by integrating quantum particle swarm optimization with accurate, although both models are effective tools. Salimbahrami and
least squares support vector machines (QPSO-LSSVM) to predict fracture Shakeri (2021) compared the ANN model with the support vector ma-
properties of fiber-reinforced cementitious composites, further proving chine (SVM) model and found that the SVM model had better agreement
the robustness of hybrid optimization models in enhancing prediction for modeling CS of RAC. J. Duan et al. (2021) employed the imperialist
accuracy. competitive algorithm to improve the accuracy of different models,
Non-destructive testing (NDT) methods, such as ultrasonic pulse including XGBoost, ANN, SVM, and ANFIS, and found that the
velocity (UPV) and hammer rebound tests, have also been applied to ICA-XGBoost model was the most accurate in modeling CS of RAC. Quan
predict mechanical properties of concrete. Revilla-Cuesta et al. (2022a); Tran et al. (2022) utilized the PSO algorithm to optimize different

2
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Table 1
Experimental tests and their statistics in the samples conducted.
Statistics Input variables Output variable
3 3 3
WA (%) WCR (ratio) FA (Kg/m ) RCA (Kg/m ) NCA (Kg/m ) WTMR (ratio) CS of RAC (Mpa)

Min 0.34 0.31 278.00 0.00 0.00 0.03 15.01

Mean 9.39 0.52 731.60 620.80 458.00 0.10 40.22
Standard deviation 4.73 0.09 187.80 300.27 304.01 0.03 12.38
Max 21.60 0.74 1212.00 1324.00 1340.00 0.20 73.08

machine learning algorithms for modeling CS of RAC and found that the coarse aggregate had a density of 1324 kg/m3, while the natural coarse
PSO-GBM model was the most accurate. aggregate had a density of 1340 kg/m3. Fine aggregate was used at
The limitations of current methods, such as uncertainties in con- densities ranging from 278 kg/m3 to 1212 kg/m3.
version models between destructive and non-destructive testing, were All materials were tested for quality assurance, and water absorption
addressed by Ali-Benyahia et al. (2019), reinforcing the need for more for the RCA was measured to be between 0.34% and 21.60%, ensuring
reliable prediction models, which this study aims to address with hybrid that the mix adhered to the design specifications.
metaheuristic-optimized models. This paper, therefore, introduced the
cascade forward neural network (CFNN) for modelling CS of RAC, and 3.2. Details of experimental design
this is the first time the CFNN model is used for this purpose. To improve
the accuracy of the CFNN model, three metaheuristic algorithms, In traditional concrete mix design, the water to binder ratio (W/B) is
namely elite single genetic algorithm (ESGA), artificial bee colony commonly used to control workability and strength. However, in the
(ABC), and PSO were used, abbreviated as ESGA-CFNN, ABC-CFNN and case of recycled aggregate concrete (RAC), the properties of both the
PSO-CFNN models. binder and aggregates play a crucial role in determining mechanical
These models are tested using practical engineering standards and behavior. Therefore, this study uses the water to total material ratio
the novelty and superiority of the proposed method is demonstrated. (WTMR), which takes into account the combined weight of the binder
This paper aims to contribute to the field by providing highly accurate and aggregates, including recycled coarse aggregates (RCA). The use of
and reliable predictions for CS of RAC and optimized construction WTMR provides a more comprehensive understanding of the water’s
practices for sustainability and efficiency. The novelty of this research influence on both the binder matrix and the interaction with aggregates,
lies in the integration of advanced hybrid metaheuristic algorithms with leading to more accurate predictions of RAC compressive strength.
CFNN, which significantly improves the prediction accuracy for RAC After reviewing the literature and materials description, several
compressive strength. Compared to traditional methodologies, these factors affect the CS of RAC including the water to cement ratio (WCR),
models minimize overfitting and computational costs while providing water absorption (WA), recycled coarse aggregate density (RCA), fine
more accurate predictions. aggregate density (FA), naturally occurring coarse aggregate density
In terms of sustainability, RAC offers significant environmental (NCA) and WTMR (Montero and Laserna, 2017; Thomas et al., 2018;
benefits by reducing the demand for natural aggregates and minimizing Plaza et al., 2021) were studied to predict the CS of RAC. An experi-
construction waste. The recycling process diverts construction debris mental plan was designed to study these parameters on the CS of RAC.
from landfills, promoting the circular economy and reducing the envi- So 272 RAC experiments were conducted to measure its CS, focusing on
ronmental impact of concrete production. However, the use of RAC can the above mentioned parameters.
also lead to challenges related to waste streams, particularly in terms of The concrete specimens were cast in cylindrical molds of 150 mm
managing residual contaminants or uneven quality in the recycled ma- diameter and 300 mm height, following ASTM C39 standards. For each
terials. Despite these challenges, the proposed model facilitates opti- mix ratio, three parallel specimens were prepared to ensure the reli-
mized mix designs that balance the mechanical performance of RAC ability of the results. All specimens were cured in water at 23 ± 2 ◦ C for
with its environmental benefits, ultimately contributing to more sus- 28 days. The compressive strength of the RAC was measured after 28
tainable construction practices. days using a universal testing machine, as per the ASTM C39 procedures.
Different mixing ratios and parameters were used in laboratory ex-
3. Recycled concrete experiments periments with water absorption of recycled aggregates ranging from
0.34% to 21.60%; Water to cement ratio in the concrete mix ranging
3.1. Description of materials used from 0.31 to 0.74; Fine aggregate density in the mix ranging from 278
kg/m3 to 1212 kg/m3; Recycled coarse aggregate density ranging from
Recycled concrete aggregates (RCA) are a key component in recycled 0 kg/m3 to 1324 kg/m3; Naturally occurring coarse aggregate density
aggregate concrete (RAC) and a sustainable alternative to traditional ranging from 0 kg/m3 to 1340 kg/m3; Water to total material ratio in the
construction materials. These aggregates come from the processing of mix ranging from 0.03 to 0.20.
demolished concrete structures and other construction waste, a vital The minimum compressive strength recorded was 15.01 MPa, and
resource to reduce environmental impact and conserve natural re- the maximum strength achieved was 73.08 MPa. The factorial experi-
sources. The process of getting RCA is involved in Fig. 1. mental design allowed for the evaluation of the combined effects of the
In this study, Ordinary Portland Cement (OPC) was used as the pri- water absorption, WCR, RCA, NCA, and FA on the compressive strength
mary binder material, with a strength grade of 42.5 MPa. The chemical of RAC.
composition of the OPC was determined following standard testing In this study, a factorial design was used to vary the levels of each
procedures and confirmed to meet the requirements of ASTM C150. The parameter (independent variable) within the range. Then randomly
fine aggregate used in the experiment had a fineness modulus of 2.75, assigned the experimental conditions to ensure unbiased results and
while the coarse aggregate consisted of both RCA and natural coarse minimize the effect of confounding variables. Then multiple reps at each
aggregate (NCA). condition were conducted to account for variability and to increase the
The RCA was sourced from demolished concrete structures, and its reliability of the results. Other variables not under investigation con-
particle size distribution was carefully graded to match the requirements stant such as cement type, aggregate gradation, curing conditions and
for structural concrete applications. The natural coarse aggregate was testing procedures are also kept to isolate the effect of independent
river gravel, with a maximum particle size of 20 mm. The recycled variables. The compressive strength values of the RAC specimens under

3
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

This combination may have contributed to a higher density of the

aggregate mix, resulting in positive correlation values. In contrast, the
lack of correlation between the WCR and the WTMR suggests that these
two factors are affected by different aspects of the concrete mix, such as
the fine aggregate and recycled aggregate densities, causing them to
vary independently. This independence is consistent with variations in
the experimental setup, where diverse water absorption rates and
aggregate compositions influence the relationship between WCR and
WTMR.

4. Methodology

The methodology for this study follows a structured, data-driven

approach, leveraging hybrid metaheuristic algorithms to optimize the
prediction of the compressive strength of recycled aggregate concrete
(RAC). The integration of ESGA, PSO, and ABC algorithms with the
CFNN forms the core of the model development. The BPNN, SVM and
XGBOost models were also applied to compare with the hybrid models
ESGA-CFNN, PSO-CFNN and ABC-CFNN. This process ensures accurate
and efficient predictions, while also optimizing the mix design to sup-
port sustainable construction practices. Below, Fig. 3 presents the step-
by-step technical roadmap of this study, which outlines the key stages
involved in data collection, model development, and validation.
Fig. 2. Correlation matrix of the input variables.
4.1. CFNN
each condition were recorded along with the corresponding values of
independent parameters to total material ratio for each trial. The Demuth et al. (2009) introduced the CFNN, which is a type of
properties of the RAC datasets are in Table 1 and Fig. 2. multilayer feedforward neural network. CFNN consists of an input layer,
In Fig. 2, the Pearson correlation matrix displays the relationships an output layer, and one or more hidden layers. Each layer in the
between the input variables. It can be observed that the correlation network is connected to the next layer through two types of connections:
between RCA and NCA is positive, which could be attributed to the one from the input layer to every layer in the network and another from
mixed use of both types of aggregates in certain experimental batches. each layer to the subsequent layers. A single-hidden-layer neural
network is generally adequate for resolving the majority of classification

Fig. 3. The step-by-step process of the study.

4
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 4. Architecture of the CFNN model.

and prediction problems (Cheng et al., 2021), as shown in Fig. 4. CFNN’s The output of the jth hidden neuron now can be computed as follows:
distinctive feature, as depicted in Fig. 3, is the direct connections be- ( )
tween the input and output layers, in addition to the standard connec- Ohj = f nethj (2)
tions between the input, hidden, and output layers. By adding more
connections within the neural network, the distribution of data is where the f function is in the range of − 1 to 1.
improved, leading to a better ability for the neural network to generalize Finally, the output of the CFNN network is calculated as follows:
(Gao et al., 1996). (
L M
)
∑ ∑
In Fig. 3, the total inputs of the jth hidden neuron is calculated as yn = g wj,n .Ohj +
2
wi,n .ui
3
for 1 ≤ n ≤ N (3)
follows: j=1 i=1

M
∑ This study seeks to predict the CS of RAC by utilizing a CFNN model.
nethj = w1i,j .ui for 1 ≤ j ≤ M (1) The model will incorporate six input variables, as outlined in Table 1. As
i=1

Fig. 5. Pseudo code of the ESGA algorithm.

5
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

a result, the CFNN model will include six additional direct connections
from the input layer to the output layer, in addition to the standard
connections. This additional feature aims to provide a more accurate
insight into the relationship between the input parameters and the CS of
RAC.

4.2. Brief overview of metaheuristic algorithms

4.2.1. ESGA
I. De Falco et al. (2002) proposed the single genetic algorithm (SGA)
with an enhancent of the basic GA (Holland, 1992) based on the mu-
tation techniques, and ESGA is the elite version of the SGA. ESGA em-
ploys a genetic algorithm that only utilizes the mutation operator and
excludes crossover. This mutation operator incorporates either
frame-shift or translocation. The translocation operator is used deter-
ministically on every selected individual in order to increase the simi-
larity to natural mutations. Following translocation, the newly
generated string is subjected to the classical error copy mechanism,
resulting in a low likelihood of point mutation. The classic tournament
selection is also utilized. In the ESGA, elitism is considered. In genetic
algorithms, the practice of elitism involves transferring the
top-performing members of the present population to the subsequent
generation without modification. This approach guarantees the reten-
tion of the most optimal solutions obtained so far, preventing them from
being lost due to random fluctuations that may occur during the
evolutionary process. In ESGA, the elite individuals are identified based
on their fitness score, and they are replicated into the next generation
without mutation. This helps to maintain a high level of fitness in the
population and can lead to faster convergence towards the optimal so-
lution. The number of elite individuals that are carried over to the next
generation is a hyperparameter that needs to be specified before running
the algorithm. The pseudocode of the ESGA algorithm is presented in
Fig. 5.
While the basic GA algorithm includes only 4 hyper-parameters (i.e.,
population size, crossover rate, mutation rate, and the number of gen-
erations) that can be tuned to achieve better convergence and global
optimum, the ESGA algorithm contains 10 parameters, including: cross-
over probability, mutation probability, selection, crossover, mutation, k-
way, the proportion or count of the top performers within the elite
group, the proportion or count of the bottom performers within the elite
group, the chosen strategy, and the size of the population.

4.2.2. PSO
PSO is a metaheuristic algorithm inspired by bird flocking or fish
schooling. Developed by J Kennedy and R Eberhart (1995a,b), PSO is
used to solve optimization problems by iteratively improving candidate
solutions. In PSO a swarm of particles (solutions) move through the
search space. Each particle moves based on its own experience and the
experience of its neighbours. The movement is influenced by three main
things:
Inertia: The particle’s previous velocity, to keep it moving in the
same direction.
Fig. 6. Flowchart of the PSO algorithm.
Cognitive Component: The particle’s own best position (personal
best), to get it back to good positions it has been to before.
Social Component: The swarm’s best position (global best), to get the optimization problems by mimicking the intelligent food-searching
particle to the best positions found by the swarm. process of a bee colony. ABC has three types of bees: employed bees,
These components balance exploration (searching new areas) and onlooker bees, and scout bees, each with a specific role in the optimi-
exploitation (focusing on known good areas) of the search space, so PSO zation process:
can converge to an optimal or near optimal solution. PSO is popular Employed bees: These bees search for food sources (solutions) and
because it’s simple, easy to implement and can handle a wide range of share the information about their quality (fitness) with onlooker bees.
optimization problems in engineering, economics and other fields. Fig. 6 Each employed bee is associated with a food source.
interprets the flowchart of the PSO algorithm. Onlooker bees: These bees evaluate the information shared by
employed bees and select food sources based on their probability values,
4.2.3. ABC which are proportional to the fitness of the solutions. Onlooker bees then
ABC is a metaheuristic algorithm inspired by the foraging behavior of exploit these food sources by exploring their neighborhood for poten-
honey bees, developed by Basturk (2006). It is used to solve tially better solutions.

6
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Scout bees: These bees randomly search the entire search space for
new food sources, replace the employed bees associated with food
sources that have not been improved for a predetermined number of
iterations.
The ABC algorithm balances exploration and exploitation by using
employed and onlooker bees to search intensively the promising areas of
the search space, and scout bees to introduce diversity by searching new
areas. This makes ABC a good and flexible algorithm to solve various
optimization problems in fields such as engineering, economics and
artificial intelligence. Fig. 7 shows the flowchart of the ABC algorithm.
In summary, these algorithms have the following similarities and
differences.
Similarities:

✓ All three algorithms—ESGA, PSO, and ABC—are population-based

metaheuristic optimization techniques that focus on finding global
optima in complex search spaces. They operate by iteratively
refining a set of potential solutions to optimize the given objective
function.
✓ Each algorithm employs an exploration and exploitation mechanism.
Exploration helps in searching new areas in the solution space, while
exploitation focuses on improving existing solutions to ensure
convergence toward the optimal solution.
✓ The fitness of the population is evaluated in each iteration using a
fitness function, such as Mean Absolute Error (MAE), to minimize
errors in predicting compressive strength (CS) of recycled aggregate
concrete (RAC).
✓ These algorithms are computationally efficient and can handle non-
linear, high-dimensional problems like RAC strength prediction.

Differences:

✓ ESGA incorporates elitism, which preserves the best solutions across

generations. Unlike traditional genetic algorithms, ESGA uses
mutation-based techniques for optimization, which increases its
ability to escape local optima and converge on the global solution.
✓ PSO is based on the social behavior of particles (solutions) in the
search space, where particles adjust their positions based on their
individual experiences and the collective experience of the swarm.
PSO’s advantage is its simplicity and the balance between explora-
tion and exploitation, but it may require fine-tuning to avoid pre-
mature convergence.
✓ ABC simulates the foraging behavior of honey bees, where the search
process is guided by employed bees, onlookers, and scouts. ABC
emphasizes exploration more, especially with scout bees searching
for new regions when existing solutions stagnate. This makes ABC
robust in escaping local minima but can result in slower convergence
compared to ESGA and PSO.

4.3. Combination of CFNN and metaheuristic algorithms

The aim of this study is to investigate the CS of RAC under different

materials and conditions and to develop AI-based models to predict the
CS of RAC for quality control or mix design optimization. The CFNN is
the main model used for CS prediction. The backpropagation algorithm
is the most common algorithm used to train the CFNN model. This al-
gorithm involves calculating the error between the network’s output
and the desired output and then adjusting the network’s weights to
minimize this error (Rahman et al., 2022). Although widely used, it has
several drawbacks such as being computationally expensive, overfitting,
local minima, gradient vanishing or exploding and sensitivity to initi-
alisation (Wani et al., 2020; Alzubaidi et al., 2021). Therefore, to
overcome these drawbacks, ESGA, PSO and ABC algorithms were pro-
Fig. 7. Flowchart of the ABC algorithm. posed as alternatives to train the CFNN model (i.e. ESGA-CFNN,
PSO-CFNN, ABC-CFNN) instead of the standard backpropagation algo-
rithm, especially the contribution of ESGA-CFNN model as a new hybrid

7
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 8. Workflow chart of the ESGA-CFNN, PSO-CFNN, and ABC-CFNN models for modeling CS of RAC.

model in this study. calculate the model’s error. In subsequent iterations, the algorithms
Metaheuristic algorithms are optimization algorithms used to solve continue to fine-tune and optimize the solutions to generate new set of
complex problems where traditional optimization techniques are not weights with lower fitness value than the previous iterations. The best
effective (Yang et al., 2016). These algorithms are population-based metaheuristic-based CFNN model is obtained when the fitness value is
meaning they maintain a set of potential solutions called a population lowest, then the model is converged. Finally, the best model is used to
and iteratively improve the quality of the solutions in the population predict CS of RAC on the testing dataset and validated in practical en-
(Agushaka and Ezugwu, 2022). To train the CFNN model for CS of RAC, gineering. Fig. 8 shows the workflow chart of hybrid models ESGA-
ESGA, PSO and ABC algorithms generated random solutions based on CFNN, PSO-CFNN and ABC-CFNN for CS of RAC and how they work.
the initial populations and number of iterations. Each iteration gener- A CFNN model with one hidden layer and 16 neurons was used in this
ates one solution per population. Therefore, a large number of solutions study. CFNN differs from traditional feedforward networks by including
will be generated to deal with the CS of RAC problem with each solution direct connections from input to each layer, which helps speed up
contains a set of weights of the CFNN model which are optimized by the convergence during training. The number of hidden neurons was
optimization mechanism of each metaheuristic algorithm (e.g. global selected after testing values between 8 and 32 neurons, with 16 neurons
optimum). yielding the best performance during cross-validation. The Adam opti-
Once the metaheuristic algorithms have generated solutions and mizer with a learning rate of 0.001 was applied to ensure stable training,
optimized the set of weights based on their behavior, these weights are and the model was trained for 500 epochs with a batch size of 10, values
then applied to the CFNN network to calculate the output (CS of RAC) that were selected based on preliminary experiments balancing
using Eq. (3) and the errors of the predictive models. During the opti- convergence speed and prediction accuracy.
mization process, a fitness function (e.g. MAE, MSE or RMSE) is used to The ReLU activation function was utilized in the hidden layer of the

8
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 9. Convergence performance of the models in modeling RAC.

CFNN. ReLU is particularly advantageous for CFNNs as it helps to The experiments were implemented using Python 3.8. The machine
mitigate the vanishing gradient problem, allowing faster and more learning models, including CFNN, were developed using TensorFlow
efficient training. The direct connections from the input to subsequent and optimized using Scikit-learn for grid search and cross-validation.
layers further enhance the model’s ability to learn complex non-linear The experiments were conducted on a Windows 10 machine equipped
relationships. The output layer used a linear activation function to with an Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50 GHz 2.50 GHz (2
predict the continuous target variable of compressive strength. processors) and 64 GB RAM. The optimization algorithms ESGA, PSO,
For the ESGA, PSO, and ABC algorithms: The ESGA, PSO, and ABC and ABC were implemented using custom Python code. Model perfor-
algorithms were configured using the following hyperparameters. For mance evaluation and visualization were carried out using Matplotlib.
ESGA, a population size of 200 was used, a mutation probability of 0.2, TensorFlow was used to implement the CFNN model, while Scikit-
and an elitism rate of 0.05. These parameters were selected based on learn was employed for grid search cross-validation and traditional
previous research (Ivan De Falco et al., 2014) and through experimen- machine learning models. Additionally, Matplotlib was used for data
tation to achieve a balance between exploration and exploitation. For visualization, and the ESGA, PSO, and ABC algorithms were imple-
PSO, a swarm size of 200 was used, an inertia weight of 0.7, a cognitive mented using custom Python scripts.
coefficient of 1.5, and a social coefficient of 1.5, based on the recom-
mendations by James Kennedy and Russell Eberhart (1995). For ABC, a 5. Results and discussions
colony size of 200 and employed bees equal was used to half the colony
size, based on recommendations by Karaboga and Basturk (2007). All 5.1. Proposed hybrid models’ results
metaheuristic algorithms ran for 500 iterations, a number that ensured
model convergence without overfitting. After the dataset was properly prepared and the workflow chart for

9
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Table 2
Predicted CS of RAC on the testing samples.
No. Measured Predicted CS by Predicted CS by Predicted CS by No. Measured Predicted CS by Predicted CS by Predicted CS by
CS (Mpa) the ESGA-CFNN the PSO-CFNN the ABC-CFNN CS (Mpa) the ESGA-CFNN the PSO-CFNN the ABC-CFNN
model (Mpa) model (Mpa) model (Mpa) model (Mpa) model (Mpa) model (Mpa)

1 26.63 28.61 27.85 29.05 29 48.5 48.12 47.71 48.02

2 38.59 36.17 37.10 39.63 30 37.11 36.87 36.14 37.39
3 35.45 35.44 34.19 34.24 31 36.31 35.51 38.46 36.72
4 26.53 26.87 26.82 28.31 32 44.92 45.36 44.98 44.62
5 43.15 43.13 44.09 43.68 33 41.11 40.83 43.43 38.59
6 52.44 49.37 51.44 51.31 34 57.25 57.03 56.78 57.15
7 19.53 20.28 20.29 20.29 35 42.64 43.13 42.54 43.66
8 25.79 25.26 27.71 26.40 36 31.72 32.62 31.09 29.23
9 26.24 26.86 26.10 26.76 37 46.63 45.85 47.53 47.07
10 46.33 46.40 46.63 46.61 38 64.71 65.92 66.69 64.69
11 20.69 20.88 20.96 22.07 39 32.67 32.38 32.58 32.68
12 62.94 64.63 64.56 64.41 40 31.73 32.05 32.43 32.63
13 48.82 48.36 49.78 49.32 41 52.58 51.46 54.84 53.25
14 43.83 42.77 43.96 42.75 42 39.69 38.80 42.63 39.69
15 37.9 37.50 37.30 37.91 43 61.92 61.94 61.20 59.83
16 57.84 58.07 61.35 59.09 44 21.65 20.27 22.39 23.43
17 39.89 39.89 39.03 40.09 45 59.78 62.26 62.05 61.38
18 45.73 46.00 45.20 46.28 46 20.15 19.11 18.03 20.67
19 58.39 61.66 62.01 61.26 47 56.93 56.18 60.32 56.73
20 24.09 25.15 25.51 24.83 48 39.2 39.37 39.11 38.69
21 52.55 52.46 52.95 52.97 49 35.96 34.96 34.01 33.34
22 32.26 33.00 33.53 31.94 50 55.19 55.42 54.55 49.98
23 38.45 35.77 37.81 35.57 51 54.3 54.47 54.81 53.83
24 24.07 24.12 23.58 25.04 52 27.3 28.87 27.54 28.51
25 30.91 28.86 31.87 31.25 53 50.87 52.33 51.81 50.10
26 33.98 33.68 32.66 31.91 54 28.28 28.44 28.44 28.69
27 34.1 33.28 33.13 34.30 55 57.18 57.74 61.11 59.24
28 37.6 38.02 40.69 38.88 – – – – –

Fig. 10. Comparison of the predicted CS of RAC by the ESGA-CFNN, PSO-CFNN, and ABC-CFNN models on the testing dataset.

10
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 11. Training performance of the BPNN, SVM and XGBoost models

the ESGA-CFNN, PSO-CFNN, and ABC-CFNN models was proposed for process for the three models, showing high convergence rates and
modeling CS of RAC, the predictive models were developed for this similar training and testing loss curves. Although the ABC-CFNN model
study. Initially, the dataset was split into two parts, with 80% of the displayed a lower convergence rate compared to the ESGA-CFNN and
entire dataset used to train the models and the remaining 20% used to PSO-CFNN models for modeling CS of RAC, it was still satisfactory in this
test their performance. It is worth noting that prior to model training, study. This suggests that the models are well-trained, free from over-
the dataset was normalized using MinMax scaling method. Next, a CFNN fitting, and capable of accurately modeling unseen data in the testing
model structure was defined with a single hidden layer and 16 hidden dataset. The optimized ESGA-CFNN, PSO-CFNN, and ABC-CFNN models
neurons selected for this network. Then, the parameters for the ESGA, were used to predict the testing dataset, and their good performance is
PSO, and ABC algorithms were established to optimize the designed demonstrated in Table 2 and Fig. 10.
CFNN model. Table 2 and Fig. 10 illustrate the high accuracy of all developed
To assess the quality of the solutions generated by the ESGA-CFNN, models in modeling the CS of RAC on the unseen datasets (i.e., testing
PSO-CFNN, and ABC-CFNN models, the fitness function MAE was uti- dataset). Among the models, the ESGA-CFNN model produced pre-
lized, and to prevent overfitting, the approach of k-folds cross-validation dictions closest to the measured CS of RAC, although the differences
was utilized, where k was equal to 5. Fig. 9 displays the optimization between the predicted values of the developed models are not

11
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 11. (continued).

n
∑
Table 3 (Predicted CSi − Actual CSi )2
Performance metrics of the ESGA-CFNN, PSO-CFNN and ABC-CFNN models i=1
R =1 −
2
n (5)
compared with the traditional machine learning models for modeling CS of RAC. ∑
(Actual CSi − Mean of actual CSi )2
Model Training metrics Testing metrics i=1

2 2
RMSE R a10-index RMSE R a10-index p10
a10 − index = (6)
ESGA-CFNN 1.614 0.983 0.972 1.144 0.991 1.000 n
ABC-CFNN 1.765 0.980 0.945 1.434 0.987 0.982
PSO-CFNN 1.719 0.981 0.945 1.561 0.984 0.982 where the dataset contains n RAC samples, and p10 represents the
BPNN 3.756 0.908 0.788 3.591 0.916 0.764 number of samples where the experimental rate value to the estimated
SVM 1.971 0.975 0.926 2.294 0.966 0.855
value ratio is between 0.90 and 1.10.
XGBoost 3.339 0.927 0.788 3.220 0.932 0.800
First, let’s look at the training metrics in Table 3. All three proposed
hybrid models do great on the training set with high R2 (above 0.98) and
significant. low RMSE (below 1.8) and a10-index (above 0.94) which measures the
accuracy of the models in modeling the top 10% of the target variable (i.
5.2. Comparison with BPNN and traditional algorithms e. CS of RAC). So all models have learned the training data well and can
predict CS of RAC.
In this study, the performance of the proposed ESGA-CFNN, PSO- Now let’s look at the testing metrics in Table 3. The proposed hybrid
CFNN, and ABC-CFNN models was evaluated by comparing them with models perform slightly worse on the testing set which is expected. But it
traditional machine learning algorithms such as Backpropagation Neu- can be seen that all the proposed hybrid models have low RMSE
ral Networks (BPNN), Support Vector Machines (SVM), and XGBoost. (1.1–1.6), high R2 (0.98–0.99) and high a10-index (0.98–1.0). So the
This comparison was performed to assess the advantages of the proposed proposed hybrid models are not overfitting and can generalize well to
models over conventional techniques in predicting the compressive new data for CS of RAC.
strength of RAC. It is worth noting that the BPNN, SVM and XGBoost From the proposed hybrid models it can be seen that ESGA-CFNN has
models (Zhou et al., 2021) used the same training and testing datasets, the lowest RMSE and highest R2 on both training and testing set and
as used for the proposed hybrid models based on the CFNN. highest a10-index on testing set. So we can say ESGA-CFNN is the best
For the BPNN model, a neural network with one hidden layer of 16 model among the three proposed hybrid models.
neurons and a ReLU activation function was employed. The model was The results of the comparison, summarized in Table 3, clearly show
optimized using the Adam optimizer, and the performance was evalu- that the ESGA-CFNN, PSO-CFNN, and ABC-CFNN models outperform
ated in terms of Mean Absolute Error (MAE). Similarly, for the SVM and the traditional algorithms (i.e., BPNN, SVM and XGBoost). The BPNN
XGBoost models, a hyperparameter search using GridSearchCV was model, while able to achieve reasonable accuracy, suffers from issues
conducted to identify the optimal parameters, ensuring a fair compari- such as overfitting and a tendency to converge to local minima, partic-
son with the proposed metaheuristic-based models. The development of ularly when dealing with complex data distributions. This limitation is
the BPNN, SVM and XGBoost models is shown in Fig. 11. evident in the slightly higher MAE values when compared to the pro-
To quantitatively assess the performance of the models, three per- posed models.
formance metrics were used, namely RMSE, R2, and a10-index (defined XGBoost and SVM, two widely used traditional machine learning
in equations (4)–(6)), as presented in Table 3. These metrics provide a algorithms, also displayed limitations when applied to this dataset.
reliable evaluation of the models’ accuracy and detect the occurrence of While XGBoost, a more advanced gradient boosting technique achieved
overfitting. moderate accuracy, it struggled with larger datasets and non-linear re-
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ lationships between variables. On the other hand, SVM with an RBF
√∑
√n kernel, performed better than XGBoost and BPNN, but still fell short
√ (Predicted CSi − Actual CSi )2 compared to the ESGA-CFNN, PSO-CFNN, and ABC-CFNN models.
√
RMSE = i=1 (4) The superior performance of the proposed models can be attributed
n
to their ability to optimize the CFNN structure more effectively by

12
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 12. Model validation plot of measured and predicted values with a margin of error within 5%, 10%, 15% and 20% respectively (a) 5% margin of error (b) 10%
margin of error (c) 15% margin of error (d) 20% margin of error

13
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Fig. 12. (continued).

14
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Table 4 developed models for prediction. The predicted CS values were

Properties of validation samples. compared with the experimentally measured CS values to evaluate the
WA WCR FA RCA NCA WTMR CS of RAC models’ accuracy. The properties of these data are described in Table 4.
(%) (ratio) (Kg/ (Kg/m3) (Kg/m3) (ratio) (Mpa) To test the accuracy of the hybrid models developed in this study for
m3) CS of RAC, six more samples were prepared. The purpose was to test
7.73 0.63 1106 72 602 0.09 36.97 ESGA-CFNN with the previous statements. The input values of these six
17.39 0.55 571 129 557 0.07 59.21 validation samples were fed into ESGA-CFNN, PSO-CFNN, ABC-CFNN,
3.01 0.55 727 615 511 0.06 23.26 BPNN, SVM and XGBoost and the predicted CS of RAC were obtained.
8.26 0.46 517 771 613 0.06 36.74
5.93 0.67 952 152 684 0.16 34.79
The results are shown in Table 5 and compared to the actual CS of RAC
5.79 0.62 963 1145 190 0.1 30.68 from experimental test.
Table 4 presents the comparison of the predicted compressive
strength (CS) values generated by six different models—ESGA-CFNN,
PSO-CFNN, ABC-CFNN, BPNN, SVM, and XGBoost—against the actual
Table 5
Comparison of predicted and actual cs for validation samples. measured CS values for six additional validation samples.
The ESGA-CFNN model consistently provides the closest predictions
No. Actual ESGA- PSO- ABC- BPNN SVM XGBoost
to the actual CS values, with small deviations across most samples. For
CS CFNN CFNN CFNN
instance, for sample 1, ESGA-CFNN predicted 37.288 MPa, which is only
1 36.97 37.288 34.769 36.082 36.420 36.812 36.582
a 0.86% deviation from the actual CS of 36.97 MPa. The PSO-CFNN and
2 59.21 60.175 61.426 60.930 59.868 57.054 53.139
3 23.26 23.136 23.074 23.753 21.673 22.071 27.845 ABC-CFNN models also performed well, but their accuracy varied
4 36.74 36.556 37.035 35.631 31.063 36.211 37.947 slightly. In sample 2, for example, PSO-CFNN predicted 61.426 MPa,
5 34.79 33.730 31.956 33.975 32.165 37.274 33.016 which overestimates the actual CS by approximately 3.75%, while ABC-
6 30.68 31.787 30.889 32.344 31.362 33.469 33.016 CFNN predicted 60.930 MPa, which is closer but still slightly over-
estimates the value. BPNN, SVM, and XGBoost show varying perfor-
avoiding local minima and overfitting, common challenges faced by mance across different samples. The deviation analysis of these models is
traditional algorithms. The metaheuristic algorithms used (ESGA, PSO, shown in Table 6. The results indicate that ESGA-CFNN offers the most
and ABC) improved the model’s generalization by efficiently exploring accurate predictions among the tested models, making it the best
the search space and identifying optimal solutions that traditional al- candidate for practical applications in predicting compressive strength
gorithms such as BPNN and SVM struggled to achieve. of RAC. PSO-CFNN and ABC-CFNN follow closely, while SVM and BPNN
In conclusion, the results validate the effectiveness of the proposed perform well in specific cases. XGBoost shows some variability, partic-
ESGA-CFNN, PSO-CFNN, and ABC-CFNN models in predicting RAC ularly in samples with lower compressive strength, suggesting that it
compressive strength. The hybrid nature of these models, combining may require further tuning to handle such cases effectively.
CFNN with advanced metaheuristic techniques, provides a significant Indeed, RAC mix design involves balancing multiple parameters such
improvement over traditional machine learning methods, making them as water-cement ratio, RCA density, fine aggregate density, and NCA
more suitable for complex engineering datasets like the one used in this density. By integrating the ESGA-CFNN model into the design phase,
study. The regression between predicted and measured CS of RAC by engineers can predict the optimal combination of these variables to
ESGA-CFNN, PSO-CFNN and ABC-CFNN is shown in Fig. 12. achieve the desired compressive strength while minimizing material
In general the results show that ESGA-CFNN is a robust and accurate usage. The model leverages the elite single genetic algorithm’s ability to
model for CS of RAC. But further analyze is need if this model is the best explore a broad solution space and identify globally optimal solutions
for the specific application (i.e. CS of RAC). for complex multi-variable problems. As a result, the ESGA-CFNN model
Fig. 12 shows that ESGA-CFNN outperformed PSO-CFNN and ABC- ensures that the design of RAC meets performance standards without
CFNN in regression level with various magrin of errors. This means excessive experimentation, thus saving time and resources. In addition,
that the relationship between measured and predicted CS of RAC is one of the major benefits of using the ESGA-CFNN model is its potential
better explained by ESGA-CFNN than others. Moreover, the results show for significant cost savings. Traditional methods for optimizing RAC mix
that ESGA-CFNN has the best fit with testing dataset based on regression designs often involve trial and error, leading to wasted materials and
analysis. Also the distance between regression line and data points in higher production costs. ESGA-CFNN eliminates much of this in-
ESGA-CFNN is better than PSO-CFNN and ABC-CFNN in modeling CS of efficiency by accurately predicting the mix ratios required to meet
RAC on testing dataset. specific performance criteria. This allows manufacturers to reduce the
To further validate the proposed ESGA-CFNN, PSO-CFNN, and ABC- amount of new material, such as virgin aggregates, while still main-
CFNN models, we prepared six additional RAC samples under different taining the required strength and durability of the concrete. Addition-
practical conditions. These samples were selected to represent a wide ally, the ability to predict performance without extensive laboratory
range of RAC properties, ensuring that the models’ performance could testing further reduces costs associated with material testing and quality
be assessed across diverse scenarios. The input parameters for these assurance processes. Also ESGA-CFNN can be used in quality control to
samples, such as water-to-cement ratio, recycled coarse aggregate den- verify that the actual CS of RAC meets the specified requirements. The
sity, and other relevant factors, were recorded and then fed into the ESGA-CFNN model provides an invaluable tool for ensuring that RAC
produced in construction projects consistently meets the required

Table 6
Deviation analysis of predicted and actual cs values.
Sample Actual CS ESGA-CFNN Deviation PSO-CFNN Deviation ABC-CFNN Deviation BPNN Deviation SVM Deviation XGBoost Deviation
(MPa) (%) (%) (%) (%) (%) (%)

1 36.97 0.86% 5.94% 2.40% 1.49% 0.32% 1.05%

2 59.21 1.63% 3.75% 2.90% 1.11% 3.65% 10.26%
3 23.26 0.53% 0.80% 2.12% 6.82% 5.11% 19.71%
4 36.74 0.49% 0.80% 3.02% 15.48% 1.44% 3.29%
5 34.79 3.05% 8.14% 2.34% 7.53% 7.15% 5.10%
6 30.68 3.61% 0.68% 5.42% 2.22% 9.12% 7.64%

15
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

compressive strength standards. By using the model as part of the quality Acknowledgements
control process, construction professionals can continuously monitor the
strength of RAC in real-time, making adjustments to the mix design if This research was supported by the State Key Laboratory of Precision
necessary. This proactive approach ensures that the concrete used in Blasting and Hubei Key Laboratory of Blasting Engineering, Jianghan
projects meets safety and performance standards, reducing the risk of University (No. PBSKL 2023A12) and Hubei (Wuhan) Explosive Tech-
structural failures or costly repairs. Furthermore, the model’s high nology Research Institute Doctoral Scientific Research Start-up Fund
predictive accuracy, as demonstrated by its superior performance met- Project (PBSKL-2024-QD-13).
rics (RMSE, R2, and a10-index), ensures that variations in RAC quality
are detected early, enabling immediate corrective actions. Overall Data availability
ESGA-CFNN can provide valuable information and aid in decision
making for RAC in construction projects. Data will be made available on request.

6. Conclusion References

This research advances the state-of-the-art in predicting RAC me- Abdollahzadeh, G., Jahani, E., Kashir, Z., 2016. Predicting of compressive strength of
recycled aggregate concrete by genetic programming. Comput. Concrete 18 (2),
chanical behavior by leveraging hybrid metaheuristic algorithms. The 155–163.
proposed models outperform conventional methods by improving pre- Agushaka, J.O., Ezugwu, A.E., 2022. Initialisation approaches for population-based
diction accuracy and generalization while maintaining computational metaheuristic algorithms: a comprehensive review. Appl. Sci. 12 (2), 896.
Akhtar, A., Sarmah, A.K., 2018. Construction and demolition waste generation and
efficiency. While the results show ESGA-CFNN outperforms in predict- properties of recycled aggregate concrete: a global perspective. J. Clean. Prod. 186,
ing RAC’s compressive strength, optimizing mix design and cost 262–281.
reduction needs further analysis and implementation. ESGA-CFNN can Ali-Benyahia, K., Sbartaï, Z.-M., Breysse, D., Ghrici, M., Kenai, S., 2019. Improvement of
nondestructive assessment of on-site concrete strength: influence of the selection
be used as a decision making tool to select the optimal mix ratio based on process of cores location on the assessment quality for single and combined NDT
target strength while minimizing high cost materials like cement. This techniques. Construct. Build. Mater. 195, 613–622. https://doi.org/10.1016/j.
can be integrated in the mix design process so engineers can try different conbuildmat.2018.10.032.
Alwash, M., Breysse, D., Sbartaï, Z.M., Szilágyi, K., Borosnyói, A., 2017. Factors affecting
RAC compositions and find the most cost effective without compro-
the reliability of assessing the concrete strength by rebound hammer and cores.
mising strength. Construct. Build. Mater. 140, 354–363. https://doi.org/10.1016/j.
Also, real time predictions from ESGA-CFNN can benefit the quality conbuildmat.2017.02.129.
control of RAC during production. By comparing predicted CS with Alzubaidi, L., Zhang, J., Humaidi, A.J., Al-Dujaili, A., Duan, Y., Al-Shamma, O., et al.,
2021. Review of deep learning: concepts, CNN architectures, challenges,
measured CS from samples taken during construction, any deviation applications, future directions. Journal of big Data 8, 1–74.
from the desired mix quality can be detected and adjustments can be Basturk, B., 2006. An artificial bee colony (ABC) algorithm for numeric function
made in the production process. While the model shows promise in optimization. In: IEEE Swarm Intelligence Symposium. Indianapolis, IN, USA, 2006.
Boussahoua, Y., Kenai, S., Sbartai, Z.M., Breysse, D., Ali-Benyahia, K., 2023. Influence of
reducing material use and maintaining quality, more studies are needed the number of cores on concrete strength assessment by nondestructive tests in old
to quantify the exact cost savings and quality control benefits in actual existing structures. Asian Journal of Civil Engineering 24 (6), 1731–1745. https://
projects. doi.org/10.1007/s42107-023-00599-0.
Cheng, S., Wu, Y., Li, Y., Yao, F., Min, F., 2021. TWD-SFNN: three-way decisions with a
Additionally, while this study focused on key variables such as water- single hidden layer feedforward neural network. Inf. Sci. 579, 15–32.
to-total-material ratio (WTMR) and recycled coarse aggregate (RCA) de Andrade Salgado, F., de Andrade Silva, F., 2022. Recycled aggregates from
content, we acknowledge that cement strength and consumption are construction and demolition waste towards an application on structural concrete: a
review. J. Build. Eng. 52, 104452.
significant factors that influence the compressive strength and durability De Falco, I., Della Cioppa, A., Maisto, D., Scafuri, U., Tarantino, E., 2014. An adaptive
of both ordinary and recycled concrete. Future studies will incorporate invasion-based model for distributed differential evolution. Inf. Sci. 278, 653–672.
these variables into the model to improve its predictive power and De Falco, I., Della Cioppa, A., Tarantino, E., 2002. Mutation-based genetic algorithm:
performance evaluation. Appl. Soft Comput. 1 (4), 285–299. https://doi.org/
applicability across a broader range of concrete types. Including cement-
10.1016/S1568-4946(02)00021-2.
related variables will enhance the model’s ability to optimize concrete Demuth, H., Beale, M., Hagan, M., 2009. Neural Network Toolbox 6 User’s Guide’. The
formulations and better reflect the full scope of influencing factors, MathWorks. Inc, MA, US.
thereby increasing its value in practical engineering applications. Deng, F., He, Y., Zhou, S., Yu, Y., Cheng, H., Wu, X., 2018. Compressive strength
prediction of recycled concrete based on deep learning. Construct. Build. Mater. 175,
In short, ESGA-CFNN is a tool for predicting compressive strength, 562–569.
optimizing mix design, and monitoring RAC quality. However, cost Deshpande, N., Londhe, S., Kulkarni, S., 2014. Modeling compressive strength of
reduction and quality optimization need to be validated in actual pro- recycled aggregate concrete by artificial neural network, model tree and non-linear
regression. International Journal of Sustainable Built Environment 3 (2), 187–198.
jects, and future research will further refine the model by including https://doi.org/10.1016/j.ijsbe.2014.12.002.
additional factors such as cement strength and consumption content. Duan, J., Asteris, P.G., Nguyen, H., Bui, X.-N., Moayedi, H., 2021. A novel artificial
intelligence technique to predict compressive strength of recycled aggregate
concrete using ICA-XGBoost model. Eng. Comput. 37 (4), 3329–3346. https://doi.
CRediT authorship contribution statement org/10.1007/s00366-020-01003-0.
Duan, Z.H., Kou, S.C., Poon, C.S., 2013. Prediction of compressive strength of recycled
Xiaoguang Zhou: Writing – review & editing, Visualization, Vali- aggregate concrete using artificial neural networks. Construct. Build. Mater. 40,
1200–1206. https://doi.org/10.1016/j.conbuildmat.2012.04.063.
dation, Software, Methodology, Formal analysis. Jian Zhou: Writing – Gao, X., Gao, X.-M., Ovaska, S., 1996. A modified Elman neural network model with
review & editing, Visualization, Validation, Methodology, Investigation, application to dynamical systems identification. In: 1996 IEEE International
Formal analysis. Joy P. Ohl: Writing – review & editing, Writing – Conference on Systems, Man and Cybernetics. Information Intelligence and Systems
(Cat. No. 96CH35929), vol. 2. IEEE, pp. 1376–1381.
original draft, Visualization, Validation, Software, Methodology,
Hammoudi, A., Moussaceb, K., Belebchouche, C., Dahmoune, F., 2019. Comparison of
Investigation, Data curation, Conceptualization. artificial neural network (ANN) and response surface methodology (RSM) prediction
in compressive strength of recycled concrete aggregates. Construct. Build. Mater.
Declaration of Competing Interest 209, 425–436. https://doi.org/10.1016/j.conbuildmat.2019.03.119.
Holland, J.H., 1992. Genetic algorithms. Sci. Am. 267 (1), 66–73.
Kabirifar, K., Mojtahedi, M., Wang, C., Tam, V.W., 2020. Construction and demolition
The authors declare that they have no known competing financial waste management contributing factors coupled with reduce, reuse, and recycle
interests or personal relationships that could have appeared to influence strategies for effective waste management: a review. J. Clean. Prod. 263, 121265.
Karaboga, D., Basturk, B., 2007. A powerful and efficient algorithm for numerical
the work reported in this paper. function optimization: artificial bee colony (ABC) algorithm. J. Global Optim. 39,
459–471.

16
X. Zhou et al. Journal of Environmental Management 371 (2024) 123068

Kennedy, J., Eberhart, R., 1995a. Particle swarm optimization. In: Proceedings of Serres, N., Braymand, S., Feugeas, F., 2016. Environmental evaluation of concrete made
ICNN’95-international Conference on Neural Networks, vol. 4. IEEE, pp. 1942–1948. from recycled concrete aggregate implementing life cycle assessment. J. Build. Eng.
Kennedy, J., Eberhart, R., 1995b. Particle swarm optimization (PSO). In: Proc. IEEE 5, 24–33. https://doi.org/10.1016/j.jobe.2015.11.004.
International Conference on Neural Networks, pp. 1942–1948. Perth, Australia. Song, Y., Wang, X., Li, H., He, Y., Zhang, Z., Huang, J., 2022. Mixture optimization of
Liao, T.W., Li, G., 2020. Metaheuristic-based inverse design of materials–A survey. cementitious materials using machine learning and metaheuristic algorithms: state
J. Materiom. 6 (2), 414–430. of the art and future prospects. Materials 15 (21), 7830.
Liu, T., Zhang, P., Cui, G., Yue, X., 2021. Fracture performance prediction of polyvinyl Su, Y., Si, H., Chen, J., Wu, G., 2020. Promoting the sustainable development of the
alcohol fiber-reinforced cementitious composites containing nano-SiO2 using least- recycling market of construction and demolition waste: a stakeholder game
squares support vector machine optimized with quantum-behaved particle swarm perspective. J. Clean. Prod. 277, 122281.
optimization algorithm. Theor. Appl. Fract. Mech. 115, 103074. https://doi.org/ Tam, V.W., Soomro, M., Evangelista, A.C.J., 2018. A review of recycled aggregate in
10.1016/j.tafmec.2021.103074. concrete applications (2000–2017). Construct. Build. Mater. 172, 272–292.
Montero, J., Laserna, S., 2017. Influence of effective mixing water in recycled concrete. Tang, Z., Li, W., Tam, V.W., Xue, C., 2020. Advanced progress in recycling municipal and
Construct. Build. Mater. 132, 343–352. construction solid wastes for manufacturing sustainable construction materials.
Nedeljković, M., Visser, J., Šavija, B., Valcke, S., Schlangen, E., 2021. Use of fine recycled Resour. Conserv. Recycl. X (6), 100036.
concrete aggregates in concrete: a critical review. J. Build. Eng. 38, 102196. Thomas, J., Thaickavil, N.N., Wilson, P., 2018. Strength and durability of concrete
Nguyen, H., Bui, X.-N., Bui, H.-B., Mai, N.-L., 2020. A comparative study of artificial containing recycled concrete aggregates. J. Build. Eng. 19, 349–365.
neural networks in predicting blast-induced air-blast overpressure at Deo Nai open- Wang, H.-Y., Hsiao, D.-H., Wang, S.-Y., 2012. Properties of recycled green building
pit coal mine. Vietnam Neural Comput. Appl. 32 (8), 3939–3955. materials applied in lightweight aggregate concrete. Computers and Concrete, An
Nguyen, L.Q., Bui, L.K., Cao, C.X., Bui, X.-N., Nguyen, H., Nguyen, V.-D., et al., 2024. International Journal 10 (2), 95–104.
Application of artificial neural networks and UAV-based air quality monitoring Wani, M.A., Bhat, F.A., Afzal, S., Khan, A.I., 2020. Advances in Deep Learning. Springer.
sensors for simulating dust emission in quarries. In: Applications of Artificial Yang, X.-S., Bekdaş, G., Nigdeli, S.M., 2016. Metaheuristics and Optimization in Civil
Intelligence in Mining, Geotechnical and Geoengineering. Elsevier, pp. 7–22. Engineering. Springer.
Piyaphipat, S., Phungpaingam, B., Musiket, K., Xi, Y., 2023. Computational viscoelastic Yonar, A., Yonar, H., 2023. Modeling air pollution by integrating ANFIS and
modeling of strain rate effect on recycled aggregate concrete. Comput. Concr. 32 (4), metaheuristic algorithms. Model. Earth Syst. Environ. 9 (2), 1621–1631.
383–392. https://doi.org/10.12989/cac.2023.32.4.383. Zhang, P., Gao, Z., Wang, J., Guo, J., Wang, T., 2022. Influencing factors analysis and
Plaza, P., del Bosque, I.S., Frías, M., de Rojas, M.S., Medina, C., 2021. Use of recycled optimized prediction model for rheology and flowability of nano-SiO2 and PVA fiber
coarse and fine aggregates in structural eco-concretes. Physical and mechanical reinforced alkali-activated composites. J. Clean. Prod. 366, 132988. https://doi.org/
properties and CO2 emissions. Construct. Build. Mater. 285, 122926. 10.1016/j.jclepro.2022.132988.
Quan Tran, V., Quoc Dang, V., Si Ho, L., 2022. Evaluating compressive strength of Zhang, P., Mao, Y., Yuan, W., Zheng, J., Hu, S., Wang, K., 2024. A critical review on
concrete made with recycled concrete aggregates using machine learning approach. modeling and prediction on properties of fresh and hardened geopolymer
Construct. Build. Mater. 323, 126578. https://doi.org/10.1016/j. composites. J. Build. Eng. 88, 109184. https://doi.org/10.1016/j.jobe.2024.109184.
conbuildmat.2022.126578. Zhang, X., Zhang, P., Yuan, W., Hu, S., 2024. Durability prediction of geopolymer mortar
Rahman, M.M., Shakeri, M., Khatun, F., Tiong, S.K., Alkahtani, A.A., Samsudin, N.A., reinforced with nanoparticles and PVA fiber using particle swarm optimized BP
et al., 2022. A comprehensive study and performance analysis of deep neural neural network, 13 (1). https://doi.org/10.1515/ntrev-2023-0214.
network-based approaches in wind time-series forecasting. Journal of Reliable Zhou, J., Huang, S., Qiu, Y., 2022. Optimization of random forest through the use of
Intelligent Environments 1–18. MVO, GWO and MFO in evaluating the stability of underground entry-type
Revilla-Cuesta, V., Ortega-López, V., Faleschini, F., Espinosa, A.B., Serrano-López, R., excavations. Tunnel. Undergr. Space Technol. 124, 104494.
2022a. Hammer rebound index as an overall-mechanical-quality indicator of self- Zhou, J., Qiu, Y., Khandelwal, M., Zhu, S., Zhang, X., 2021. Developing a hybrid model of
compacting concrete containing recycled concrete aggregate. Construct. Build. Jaya algorithm-based extreme gradient boosting machine to estimate blast-induced
Mater. 347, 128549. https://doi.org/10.1016/j.conbuildmat.2022.128549. ground vibrations. Inter. J. Rock Mech. Mining Sci. 145, 104856.
Revilla-Cuesta, V., Shi, J.-y., Skaf, M., Ortega-López, V., Manso, J.M., 2022b. Non- Zhou, J., Shen, X., Qiu, Y., Shi, X., Du, K., 2023. Microseismic location in hardrock metal
destructive density-corrected estimation of the elastic modulus of slag-cement self- mines by machine learning models based on hyperparameter optimization using
compacting concrete containing recycled aggregate. Developments in the Built bayesian optimizer. Rock Mech. Rock Eng. 56, 8771–8788.
Environment 12, 100097. https://doi.org/10.1016/j.dibe.2022.100097. Zhou, J., Shen, X., Qiu, Y., Shi, X., Khandelwal, M., 2022. Cross-correlation stacking-
Salimbahrami, S.R., Shakeri, R., 2021. Experimental investigation and comparative based microseismic source location using three metaheuristic optimization
machine-learning prediction of compressive strength of recycled aggregate concrete. algorithms. Tunnel. Undergr. Space Technol. 126, 104570.
Soft Comput. 25 (2), 919–932. https://doi.org/10.1007/s00500-021-05571-1. Zhu, X., Chen, X., Shen, N., Tian, H., Fan, X., Lu, J., 2018. Mechanical properties of
pervious concrete with recycled aggregate. Comput. Concr. 21 (6), 623–635.

17