Muhammad et al.

AI in Civil Engineering (2024) 3:21 AI in Civil Engineering

https://doi.org/10.1007/s43503-024-00040-8

ORIGINAL ARTICLE Open Access

An improved prediction
of high‑performance concrete compressive
strength using ensemble models and neural
networks
Umar Jibrin Muhammad1, Ismail I. Aminu2, Ismail A. Mahmoud3, U. U. Aliyu4, A. G. Usman5,6,
Mahmud M. Jibril7* , Salim Idris Malami2,7* and Sani I. Abba8

Abstract
Traditional methods for proportioning of high-performance concrete (HPC) have certain shortcomings, such as high
costs, usage constraints, and nonlinear relationships. Implementing a strategy to optimize the mixtures of HPC can
minimize design expenses, time spent, and material wastage in the construction sector. Due to HPC’s exceptional
qualities, such as high strength (HS), fluidity and resilience, it has been broadly used in construction projects. In
this study, we employed Generalized Regression Neural Network (GRNN), Nonlinear AutoRegressive with exog-
enous inputs (NARX neural network), and Random Forest (RF) models to estimate the Compressive Strength (CS)
of HPC in the first scenario. In contrast, the second scenario involved the development of an ensemble model using
the Radial Basis Function Neural Network (RBFNN) to detect inferior performance of standalone model combinations.
The output variable was the 28 Days CS in MPa, while the input variables included slump (S), water-binder ratio (W/B)
%, water content (W) kg/m3, fine aggregate ratio (S/a) %, silica fume (SF)%, and superplasticizer (SP) kg/m3. An RF
model was developed by using R Studio; GRNN and NARX-NN models were developed by using the MATLAB 2019a
toolkit; and the pre- and post-processing of data was carried out by using E-Views 12.0. The results indicate that in the
first scenario, the Combination M1 of the RF model outperformed other models, with greater prediction accuracy,
yielding a PCC of 0.854 and MAPE of 4.349 during the calibration phase. In the second scenario, the ensemble of RF
models surpassed all other models, achieving a PCC of 0.961 and MAPE of 0.952 during the calibration phase. Overall,
the proposed models demonstrate significant value in predicting the CS of HPC.
Keyword High-performance concrete; Generalized Regression Neural Network; NARX neural network; Random
Forest (RF)

*Correspondence:
Mahmud M. Jibril
[email protected]
Salim Idris Malami
[email protected]
Full list of author information is available at the end of the article

© The Author(s) 2024. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 2 of 18

1 Introduction Nevertheless, the regular impertinent experimental tech-

Concrete is widely hailed as a prominent construction nique to determine the CS of HPC is somehow laborious,
material due to its exceptional versatility and durabil- cumbersome and costly. Since a range of situations and
ity, playing a significant role in shaping the advancement materials could affect the CS, the conventional formula-
of modern built environments. Concrete is made up of based techniques face many limitations in their predic-
a mixture of cement, aggregates, water and frequently, tive usefulness and are inadequate to offer satisfactory
extra additives. High compressive strength (HCS), resist- achievements. Most conventional techniques for fore-
ance to fire and weathering, and adaptability to various casting concrete strength rely on statistical analysis as
shapes and forms are only a few of its amazing qualities well as linear or nonlinear regression models. However,
(Malami et al., 2018). As a new type of advanced con- it is challenging to obtain an exact equation, because it
crete, high-performance concrete (HPC) is made with necessitates a variety of methods and expertise. In order
traditional ingredients and methods, and has all the to overcome these challenges, machine learning (ML)
mechanical properties necessary for concrete construc- approaches have been implemented to forecast the CS of
tion, plus high levels of durability, workability, and vol- HPC. Based on the established studies in various fields,
ume stability (Liu, 2022). The process of its production including concrete construction, classical tools have
involves the use of admixtures for enhancing its prop- been used for analyzing and predicting various phenom-
erties. The admixtures added include silica fume, nano- ena. However, there are limitations in terms of accuracy,
silica, blast furnace slag and fly ash, purposed to enhance especially when dealing with complex and nonlinear
the HPC-CS (Bui et al., 2018). The induction of HPC is relationships. In recent years, the development of AI has
propelled by its superior compressive strength (CS), lead- provided more promising results in finding accurate and
ing to its widespread adoption in civil engineering for the non-linear relationships than classical regression tools,
construction of high-rise buildings, bridges, and heavily because AI is capable of capturing intricate patterns and
loaded structures. dependencies within data, thus enabling better represen-
HPC is an attractive option for a diversity of construc- tation of underlying processes.
tion applications because of its greater mechanicality and With 173 datasets, Nikoo et al. (2015) anticipated
durability. However, it has several remarkable obstacles the CS of concrete by using neural networks (NN) and
that must be overcome for its adoption. Given that lit- regression models, finding that the Artificial neural net-
tle change may deliver substantial influence on the final work (ANN) model could deliver greater accuracy than
properties of HPC, maintaining exact control over its the regression model. Vidivelli and Jayaranjini (2016)
components and mix proportions becomes one of the employed ANN to predict the HPC-CS with ten indus-
main requirements (Jibril et al., 2023a, b, c, d). Another trial waste-independent variables at diverse curing dura-
challenge is how to achieve the desired workability with- tions. Making use of a different classification scheme,
out sacrificing strength; to address this problem, super- Mustapha and Mohamed (2017) demonstrated the accu-
plasticizers and sophisticated admixtures are frequently racy of weighted support vector machine (wSVM) models
used. In addition, both considerable heat created during in estimating the CS of 1,030 specimens of HPC. Moreo-
curing and increased susceptibility to shrinkage must be ver, Doddy et al. (2018) studied the HPC-CS by using a
carefully handled to prevent thermal cracking (Behnood hybrid machine learning algorithm, namely intelligence-
et al., 2017; Cheng et al., 2014; Han et al., 2019). Con- optimized support vector regression (IO-SVR), with the
cerns also arise about sustainability, because some HPC overall results showing that the hybrid IO-SVR algorithm
combinations may use certain components with large outperformed other AI methods for the analysis.
carbon footprints (Lim et al., 2004, Zhang et al., 2017; J. To predict the HPC-CS, D. Bui et al. (2018) explored
Zhang & Zhao, 2017). Overall, although HPC provides the HPC characteristics by establishing an expert sys-
high performance, it is necessary to pay rigorous atten- tem, which could determine the compressive and ten-
tion to details in its design, manufacturing and construc- sile strength (CS & TS); their study finally found that the
tion, so as to overcome the above difficulties and reach modified firefly algorithm (MFA-ANN) hybrid system
its full potential in contemporary infrastructure projects. could provide the most suitable values based on an ANN
Practical experience shows that the quality control of model together with a MFA. Yu et al. (2018) utilized SVM
concrete is highly important for safe, reliable and dura- to predict HPC CS with 1,761 datasets and enhanced cat
ble structures. An important parameter for quality con- swarm optimization, demonstrating a recognition abil-
trol of concrete is the 28-day curing of concrete, which ity and prediction accuracy superior to other models.
is often used to determine the CS of concrete. In civil Al-shamiri et al. (2019) studied the CS of high-strength
engineering practice, CS is a crucial constraint in the concrete (HSC) with extreme learning machine (ELM),
structural design of reinforced concrete structures (RCS). ANN and back propagation (BP) algorithms, proving
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 3 of 18

that ELM outperformed ANN in determining the CS. research facilitates the tracking of emerging topics,
Feng et al. (2021) implemented ANFIS, PSO, ACO and commonly researched areas, and the evolution of sci-
DEO metaheuristic algorithms to analyze self-consol- entific concepts. It also aids in identifying focal points
idating concrete characteristics for non-destructive and gaps within research landscapes, thereby assisting
tests, proving that these algorithms could be used as an researchers and policymakers. At a national level, bib-
effective alternative approach. In addition, Abisek et al. liometric evaluation provides insight into the leading
(2022) employed an Python ML tool to compare actual countries in HPC research and development, enabling
strengths, identifying that RF produced more consider- a comparative assessment of global contributions and
able results than other techniques. On the other hand, collaborations. This information is beneficial for fund-
Liu (2022) compared XGBoost, SVR and Random Forest ing agencies, governments, and academic institutions
(RF) machine learning algorithms in determining the CS to formulate their research strategies, allocate invest-
of HPC, finding that the combined XGBoost model was ments, foster international partnerships, and promote
more precise than other algorithms. Sufian et al. (2023) innovation in HPC-related fields. By linking the uti-
used Gradient Boosting (GB), SVM, and Multi-Layer Per- lization of HPC to keyword trends and geographical
ceptron algorithms to measure the flexural strength (FS) research distribution, bibliometric analysis enables a
in ultrahigh-performance concrete (UHPC), concluding comprehensive assessment of how different countries
that the GB technique was most efficient when k-fold and research areas are advancing technological fron-
cross-validation was deployed. tiers. Numerous studies have employed this analysis to
To estimate the HPC-CS, this study employed a com- spotlight the research areas, keywords, and countries
bination of Generalized Regression Neural Networks engaged in HPC utilization. For instance, Jibril et al.
(GRNN), NARX neural networks (NARX), Random (2023a), Jibril et al. (2022), and Zimit et al. (2023) have
Forest (RF), and an ensemble model using Radial Basis utilized bibliometric analysis to enhance understand-
Function Neural Networks (RBFNN). By combining sev- ing of HPC research dynamics.
eral sophisticated prediction models, this study created As indicated in the literature mentioned above, the
a novel method for resolving the CS prediction issue. A Google Scholar website reported 17,504 research docu-
broad range of input variables were used in this study, ments published from 1964 to the present that used
including slump (S), water binder ratio (W/B) %, water the viability of ML models for concrete CS prediction
content (W) kg/m3, fine aggregate ratio (S/a) %, silica of HPC. A scan of these 17,504 documents is shown in
fume (SF) %, and superplasticizer (SP) kg/m3. Such a Fig. 1, which illustrates 49,932 phrases that appeared
wide range of input parameters give consideration to sev- repeatedly across the studies, demonstrating a lack of
eral variables that may have an impact on HPC proper- general cooperation in understanding how to forecast
ties, so as to improve the precision of the CS predictions. concrete CS.
In addition, this study introduced an ensemble model as
a novel remedy for the shortcomings of solo models. This
strategy offers a creative technique to address difficulties 2 Investigational data used in this study
with modelling and significantly raises forecast accuracy. This study uses three different ML algorithms: RF, NARX
The research evaluates and compares the performance Neural Networks, and GRNN. And a thorough method-
of the various models using metrics such as the Pear- ology is proposed in this research to predict the HPC-CS
son Correlation Coefficient (PCC) and Mean Absolute by utilizing the experimental data provided by (Lim et al.,
Percentage Error (MAPE). The originality of this study 2004). The dataset collection from the study includes the
is underscored by the identification of the most effec- following input variables: slump (S), water binder ratio
tive model combinations in different scenarios, offering (W/B) %, water content (W) kg/m3, Fine aggregate ratio
valuable insights for predictive modeling in the field of (S/a) %, silica fume (SF) %, and superplasticizer (SP) kg/
concrete engineering. This investigation leverages state- m3. The output variable is the 28 Days CS in MPa (see
of-the-art technologies, utilizing specialized software Fig. 2). The goal of this strategy is to create reliable
tools for modeling and data processing, including MAT- and accurate predictive models for HPC compressive
LAB 2019a, R studio, and E-Views 12.0. This approach strength, which is a crucial variable for building engineer-
holds the potential to advance scientific research and ing. The main flow diagram for the methodologies used
engineering practices in the realm of concrete. in this research is shown Fig. 3. The recommended mod-
Bibliometric evaluation plays a crucial role in com- elling approach utilizes the conventional feature extrac-
prehending the impact and trends of research in the tion method to simulate CS based on the CC among the
field of HPC, both in terms of keywords and coun- parameters. Equation (1) gives the input combinations
tries. Through the analysis of keywords, bibliometric
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 4 of 18

Fig. 1 a the terminologies most frequently used in the literature on concrete’s HPC-CS employing ML models (1964–2023); and b the nations
that adopted HPC


used in the training of NARX NN, GRNN, and RF mod-  C1 = SF + SP
els based on sensitivity analysis. HPC − CS = C2 = W + s/a + W /B + s
 C3 = SF + SP + W + s/a + W /B + s
(1)
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 5 of 18

Fig. 2 Embedded bar chart for the input and output variables
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 6 of 18

Fig. 3 Flowchart of the proposed methodology used in this study

The numerous input sections needed to train the where "Yi" stands for the normalized measure in the
models are represented by the Basic M1, M2, and M3. dataset, "Yu" for the unnormalized amount, "Ymin" for
These sections encompassed the variables SF, SP, W, the minimum measure, and "Ymax" for the maximum
s/a, W/B, and s, along with additional related param- measure.
eters. In total, 77 datasets were gathered for this study,
which were subsequently divided into 54 (70%) calibra-
tion datasets and 23 (30%) verification datasets using 3 Theory for ML models
3.1 Nonlinear auto‑regression with eXogenous inputs
an external validation methodology. The input was then
neural network (NARX‑NN)
standardized to the range from zero to one, so as to
improve the integrity and prediction accuracy. Equa- One frequently utilized type of ANN in the realm of
tion 2 was used to remove data redundancy. time series forecasting and system identification is
the Nonlinear AutoRegressive with eXogenous inputs
Yu− Ymin NARX-NN. This particular model exhibits high per-
Yi = (2) formance in dynamic system contexts such as those
Ymax− Ymin
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 7 of 18

   
encountered in control engineering, finance, and envi- y(t), y(t − 1), . . . ., y t − ny , x(t + 1),
ronmental modeling. The fundamental tenet of NARX y(t + 1) = F
x(t), x(t − 1), . . . ., x(t − nx )
neural networks lies in their ability to represent sys- (3)
tems with intricate interdependencies by incorporat-    
ing both previous inputs and previous outputs as input y(t), y(t − 1), . . . ., 
 y t − ny , x(t + 1),
y(t + 1) = F
parameters for predicting future output states (Jibril, x(t), x(t − 1), . . . ., x(t − nx )
et al., 2023a). Due to their extraordinary effectiveness (4)
when associated with conventional NN, NARX neural where F is a neural network membership function (NN
networks are regarded as something similar to Turing MF);  y(t − 1) is the NARX output at time t for time
computers. t + 1 (which is an anticipated
  value of y for time t + 1);
An NARX-NN is composed of two distinct designs: y(t), 
 y t − ny donate NARX past output
y(t − 1), . . . ., 
open loop, also known as series–parallel, and closed  
y(t), y(t − 1), . . . ., y t − ny ; xt + 1, xt,…, xtnx donate
loop, also known as parallel (See Fig. 4a).
NARX input; n ­ x, epresenting multiple input delays, and

Fig. 4 a structure of NARX-NN; b basic structure of RF model; c GRNN Best model structure; and d RBFNN structure
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 8 of 18

­ny, representing multiple output delays, serve as exam- radial basis function’s outputs when a fresh input is sup-
ples of desired output values in time series prediction plied for prediction, based on the similarity between the
(Boussaada et al., 2018). NARX models can predict the input and the stored data points. An input point would
value of a time series using a closed-loop technique, have more impact on the prediction if it is closer to the
where the projected values are fed back into the model training data (Gholamrezaei & Ghorbanian, 2007). The
instead of the present target time series. Alternatively, in best model combination structure is shown in Fig. 4c.
the open-loop strategy, the model takes the entire time
series as input (Ouyang, 2017).
3.4 RBFNN model
3.2 RF algorithm Neural networks have proven to be advantageous in deal-
As a kind of supervised machine learning algorithm, ing with the complexity of natural systems characterized
random forest has gained much popularity, as shown in by diverse inputs, ultimately enhancing model efficiency
many literatures works. Its applications can be found in and improving result accuracy (Mayet et al., 2022). A
different research domains for its effective results (Jibril supervised training method suggested by Broomhead and
et al., 2023a, b, c, d). Due to its simple operation, RF could Lowe was used to create RBFNN, a kind of FFNN (Page
be the best way to realize physical models for large-scale et al., 1988; Pazouki et al., 2022). The fundamental struc-
hazard assessments in various catchments (Pengcheng ture of RBFNN is identical to that of MLP, with a restric-
et al., 2020). However, RF was noticed in 2001 as an tion that only one hidden layer is adopted, so RBFNN
ensemble machine learning approach for prediction by includes an input layer, a hidden layer, and an output
using a large number of classification or regression trees, layer. Among the various basic functions employed in
because of its high generality and strength. Breiman gave RBF networks, the "Gaussian" function stands out along-
the knowledge about impurity for regression trees as the side linear, polynomial, spline, and multi-degree inverse
total squared deviations between the estimated outcomes functions (Moodi et al., 2022). The rapid convergence
and the detected values. The regression trees can be rate of an RBFNN can be attributed to the Gaussian acti-
trained for some of the rules based on the response vari- vation function, which shapes the nodes of the hidden
ables, which are constructed to partition observation, as layer while the input nodes facilitate the transfer of vari-
far as the results to be predicted could get the minimum ables from the input layer to the hidden layer. This par-
level of node impurity, while the main results gathered ticular kind of neural network processes input the signals
from the trees get noticed as the final output of RF (Jibril in a manner essential to its basic operation. The hidden
et al., 2022). The basic structure of RF is shown in Fig. 4b. layer’s overall output is normally obtained by the last
layer in a straightforward linear function. The RBFNN
3.3 GRNN‑model process is described as follows:
GRNN, a sort of ANN, is generally used for regression Then, an RBF (G) computes the final of the i-th RBF
tasks, although it can also be modified for classifica- node (hi) by utilizing the equations below. First, the
tion tasks. Well renowned for being straightforward and radial separation (di) for the i-th RBF node in the hid-
effective at training, GRNN is a useful tool for regression, den layer between the input vector (x) and the Center
particularly when working on continuous data (Li et al., (ci) is determined.
2013; Zimit et al., 2023). It is a member of the Radial di = �X − CI � (5)
Basis Function Networks family, employing activation
functions with radial basis functions, which are centered
in particular data locations. Unlike conventional neural hi = G(di, σi ) (6)
networks, which need iterative training (such as gradient where X is the width of the i-th RBF node in the hidden
descent), GRNN adopts a kind of one-pass learning strat- layer, and G is the RBF function. The output of the output
egy (Hannan et al., 2010). It is computationally efficient, layer is defined as below:
because it computes the weights during training and uses
1
them immediately for predictions. The key idea of GRNN y = f (x) = Wi h i (7)
is to keep training data in the network’s hidden layer. The i=1
radial basis function centered at the input point benefits where ­wi is the weight between the i-th hidden layer node
from each training sample. During inference, the func- and the output neuron; and l is the total number of neu-
tion is employed to produce predictions (Aliyu et al., rons in the hidden layer. The basic structure used in the
2021). GRNN models compute a weighted average of the ensemble model for RBFNN is shown in Fig. 4d
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 9 of 18

3.5 Ensemble learning technique (ELT) 4 Result of the standalone model

Ensemble methods (ET) have been successfully used 4.1 Result and discussion
in a variety of sectors, including civil engineering, Data preparation, model development, and computa-
to enhance the precision of final estimates (Nourani tional insight extraction are described in this section. In
et al., 2018). In the field of ML, a technique called this study, GRNN, NARX-NN, and RF models were used
Ensemble Learning Techniques (ELT) creates several to assess the HPC-CS performance of other components.
predictors from a single model to improve the perfor- Two ML models (GRNN and NARX) were built in MAT-
mance of the final estimation. The main goal of using LAB software, and an RF algorithm was built in R Studio.
an ensemble is to get more accurate and reliable esti- E-Views 12.0 was employed for pre- and post-processing
mates than those from a single model. Homogeneous of the datasets for model construction. Through the use
and heterogeneous ensembles are two main categories of MATLAB code, the GRNN and NARX models were
for researchers to divide ensemble learning approaches trained and validated. Building GRNN and NARX-NN
(Abba et al., 2020). Employing a single, identical learn- models requires choosing a proper number of hidden
ing strategy, such as a neural network (NN), is referred nodes. This choice was made using the Eq. (2n (1/2) + m)
to as homogeneous ELT. Contrarily, heterogeneous to (2n + 1), where ’n’ denotes the number of input neu-
ELT adopts a variety of teaching strategies. A het- rons and ’m’ denotes the quantity of output nodes.
erogeneous ensemble is typically recommended for This study identifies the optimal GRNN and NARX-
attempts of making precise forecasts while addressing NN structures by using various hidden nodes, cali-
the problem of model variety (Elkiran et al., 2019). Due bration epochs, emotional hormones, and activation
to the poor performance of model combinations, an functions. Due to unpredictable conditions and the fea-
ensemble of RBFNN models was therefore used in this tures of water sources, relationships between HPC-CS
study to improve prediction accuracy; and ensembles restrictions may not be linear. The correlation between
of M1, M2, and M3 combinations were selected for the the input and output variables is shown in Fig. 5 (Cor-
GRNN, NARX-NN and RF models (RBFNN-GRNN, relation matrix). Table 2 presents the descriptive sta-
RBFNN-NARX, and RBFNN-RF). tistics of the datasets, along with essential information
necessary for model development. These statistical
metrics serve the purpose of elucidating and effectively
3.6 Performance evaluation criteria communicating the significant characteristics of the
A group of predetermined criteria and metrics were datasets, facilitating comparisons between variables or
used for evaluating models’ effectiveness, efficiency, groups, detecting anomalies, ensuring the reliability
and performance quality. The correctness of the mod- of the data, and informing decision-making processes.
els was examined in this study by using five statisti- Furthermore, these metrics aid in data exploration and
cal metrics: R, PCC, MSE, MAE, and MAPE. Table 1 enable researchers and analysts to derive meaning-
provides the formal ranges of the performance crite- ful insights from the collected data. The step-by-step
ria, which are frequently used in evaluating how well a procedure that was followed during the modelling
projected model performs. of GRNN NARX-NN is shown in Tables 3 and 4. The
calibration and verification results predicted by GRNN,

Table 1 List of the models’ performance metrics and their formulas

Name Formula Range
  
R √ N(

HPC−CSO HPC−CSP )−( HPC−CSO )( HPC−CSP )
  
(− 1 < R < 1)
[N HPC−CSo2 ](HPC−CSo)2 ][N HPC−CSp2 ]( HPC−CSp)2 ]
N   
PCC i=1 HPC−CSO −HPC−CSom HPC−CS  P
P . −HPC−CS (− ∞ < R < 1)
  2
N
[HPC−CS(P) −HPC−CSO ] 2  O
P −HPC−CS
HPC−CS
i=1

N
MSE i=1 HPC−CS(p) −HPC−CS(o) )2 (0 < MSE < ∞)
N
N
MAE |HPC−CS (p) −HPC−CS(o) |
i=1 (0 < MAE < ∞)
 N 
MAPE 100 N  HPC−CS(O) −HPC−CS(P)  (0 < MAPE < 100)
N I=1  HPC−CS(o) 

where HPC-CS(p) & HPC − CS 

P . depict the predicted HPC-CS; HPC − CSP predicts the mean HPC-S; HPC-CS(o) demonstrates HPC-CS; HPC − CSi (o) overbar depict the
observed mean HPC-CS; N depicts the number of the data set
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 10 of 18

Fig. 5 Relationship between the input and output variables

Table 2 Descriptive statistics for the input and output datasets

Parameters Mean Median Mode Standard Sample Variance Kurtosis Skewness Minimum Maximum
deviation

S 207.79 205.00 200.00 14.50 210.19 −0.92 0.22 180.00 235.00

W/B 0.12 0.00 0.00 0.46 0.21 13.34 3.83 0.00 2.00
W 0.32 0.00 0.00 4.08 16.67 10.07 −2.50 −20.00 5.00
s/a 0.05 0.00 0.00 0.86 0.73 2.73 0.16 −2.00 2.00
SF 0.00 5.00 5.00 7.65 58.55 −0.76 −0.93 −15.00 10.00
SP −0.21 0.50 0.00 4.52 20.43 0.77 −0.75 −12.50 9.90
HPC-CS −0.34 1.00 1.00 5.40 29.15 0.06 −0.49 −14.00 12.00

NARX-NN and RF are shown in Table 5. Even though It is evident in Table 5 that the GRNN model deliv-
the performance measures used in this study, including ers the highest PCC and R values, as well as the
MAPE, MAE and MSE, are unitless PCC performance lowest MSE, MAPE and MAE values, when utiliz-
measures, the results in Table 5 demonstrate that R is ing model combination 3 (M3) with PCC = 0.783,
dimensionless, as a result of the normalization in the R = 0.412, minimum MSE = 0.061, MAPE = 10.945%,
modelling technique. Table 3 shows the surprising and MAE = 0.024 during the calibration phase, which
responses of the AI-based pragmatic models to various is highlighted in bold. On the other hand, the NARX
combination models of HPC-CS (Mpa). During the cal- Neural Network Model attains the maximum PCC and
ibration phase, the predictive power was 50%–70% for R values, along with the lowest MSE, MAPE and MAE
HPC-CS (MPa), 70%–89% for PCC, and 3–5 for MAE. values, under combination 1 (M1) with PCC = 0.819,
 Muhammad et al. AI in Civil Engineering

Table 3 Step-by-step procedure for modelling of GRNN

(2024) 3:21

Model Network type Number Training algorithm Epochs Spread constant Adaption Performance Number Number of Transfer function
of inputs learning function of layers neurons
function

Calibration Phase
GRNN-M1 Generalized regression 2 Levenberg Marquardt 10 1 LEARNGDM MSE 2 4 TANSIG
GRNN-M2 Generalized regression 4 Levenberg Marquardt 10 1 LEARNGDM MSE 2 3 TANSIG
GRNN-M3 Generalized regression 6 Levenberg Marquardt 10 1 LEARNGDM MSE 2 4 TANSIG
Verification Phase
GRNN-M1 Generalized regression 2 Levenberg Marquardt 10 1 LEARNGDM MSE 2 4 TANSIG
GRNN-M2 Generalized regression 4 Levenberg Marquardt 10 1 LEARNGDM MSE 2 3 TANSIG
GRNN-M3 Generalized regression 6 Levenberg Marquardt 10 1 LEARNGDM MSE 2 4 TANSIG
Page 11 of 18
 Muhammad et al. AI in Civil Engineering

Table 4 Step-by-step procedure for modelling of MARX-NN

(2024) 3:21

Model Network type Number of Training algorithm Epochs Training function Adaption Performance Number Number of Transfer function
inputs learning function of layers neurons
function

Calibration phase
NARX-NN-M1 NARX 2 Levenberg Marquardt 10 TRAINLM LEARNGDM MSE 2 3 TANSIG
NARX-NN-M2 NARX 4 Levenberg Marquardt 10 TRAINLM LEARNGDM MSE 2 5 TANSIG
NARX-NN-M3 NARX 6 Levenberg Marquardt 10 TRAINLM LEARNGDM MSE 2 7 TANSIG
Verification Phase
NARX-NN-M1 NARX 2 Levenberg Marquardt 10 TRAINLM LEARNGDM MSE 2 3 TANSIG
NARX-NN-M2 NARX 4 Levenberg Marquardt 10 TRAINLM LEARNGDM MSE 2 5 TANSIG
NARX-NN-M3 NARX 6 Levenberg Marquardt 10 TRAINLM LEARNGDM MSE 2 7 TANSIG
Page 12 of 18
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 13 of 18

Table 5 Results of standalone models in calibration and verification

Models Calibration Phase Verification Phase
R PCC MSE MAPE MAE R PCC MSE MAPE MAE

GRNN-M1 0.399 0.806 0.007 10.913 0.025 0.334 0.566 0.004 4.523 0.018
GRNN-M2 0.131 0.526 0.008 11.872 0.027 0.090 0.499 0.005 4.898 0.020
GRNN-M3 0.412 0.783 0.006 10.945 0.024 0.340 0.580 0.004 4.526 0.018
NARX-M1 0.816 0.819 0.003 5.814 0.016 0.527 0.609 0.003 3.550 0.016
NARX-M2 0.724 0.783 0.004 5.981 0.017 0.595 0.623 0.003 3.086 0.015
NARX-M3 0.811 0.831 0.003 4.959 0.015 0.531 0.606 0.003 3.584 0.016
RF-M1 0.839 0.854 0.002 4.349 0.015 0.627 0.643 0.003 3.008 0.014
RF-M2 0.720 0.758 0.004 6.989 0.017 0.592 0.613 0.003 3.518 0.015
RF-M3 0.884 0.894 0.002 4.311 0.013 0.518 0.626 0.003 3.840 0.017

R = 0.816, minimum MSE = 0.003, MAPE = 5.814%, datasets, so they are a powerful tool in statistical evalua-
and MAE = 0.016 during the calibration phase, which tion and information visualization.
is highlighted in bold. Similarly, under combination 1
(M1), the RF model yields the highest PCC and R val- 4.2 Result of the ELT
ues, as well as the lowest MSE, MAPE and MAE values, Ensemble learning strategies were employed in this study
with PCC = 0.854, R = 0.839, minimum MSE = 0.002, to enhance the predictive accuracy and robustness of the
MAPE = 4.394%, and MAE = 0.015, which is highlighted models for estimating the CS of HPC. By combining mul-
in bold. In general, RF-M1 outperforms other models tiple models, including GRNN, NARX neural network
in the calibration phase. The aforementioned method- and RF, in conjunction with an ensemble model using
ologies are acknowledged for their effectiveness in han- RBFNN, this study could take advantage of the strengths
dling models characterized by multiple uncontrolled of every model, while mitigating their weaknesses.
parameters, minimizing the error function, and miti- Ensemble learning helps capture diverse styles and
gating data fitting issues in accordance with empirical insights from different models, leading to extra correct
evidence. These methods consistently demonstrate and reliable predictions. Additionally, this approach facil-
their suitability for addressing highly complex nonlin- itates decreasing the threat of overfitting and improves
ear scenarios. the generalization capacity of the predictive models,
Based on their distinct MAPE values, these models making them a treasured tool for optimizing HPC mix-
are graphically compared in Fig. 6. It is evident that tures in creative applications. In the second phase of this
the models’ performance deteriorates as error counts study, RBFNN (M1, M2 and M3) for both the standalone
increase, and conversely, improves with lower error model and an ensemble were used to provide more pre-
counts. In terms of performance accuracy, RF outper- cise predictions, while improving the exactness of the
forms GRNN and NARX-NN, as illustrated in Fig. 8. standalone model. Both RBFNN-RF and RBFNN-NARX-
Further elaboration on MAPE measurements will be NN achieved the maximum level of accuracy, as shown
provided below. in Table 6, ranging from 95% to 96%. Given that RBFNN
As shown in Fig. 6, the box plot, also referred to as the can handle a wider variety of input data types and set-
box-and-whisker plot, is a visual representation depicting ups and is flexible to vagaries in input data or ecologi-
the distribution of a dataset. It offers a visible summary cal variations, this result is not surprising. Using scatter
of the statistics’ vital tendency, dispersion, and skewness. plots and a response plot, Figs. 7 and 8 relate the recital
The field inside the plot represents the interquartile range of the best combination to the observed data. Scatter
(IQR), which encompasses the central 50% of the infor- plots are often employed in data analysis and scientific
mation. The line inside the field marks the median or the research, because they can help reveal spot trends, con-
center value of the dataset. The whiskers extend from the nections, and patterns in data. As shown in Fig. 8, the
edges of the box to expose the range of the information, response curve exhibits a strong agreement between
apart from outliers, which, if present, are plotted as indi- the experimental and projected values. The MAPE and
vidual factors beyond the whiskers. Box plots are treas- PBAIS data illustrate how the time series plot can dis-
ured for comparing distributions, figuring out outliers, play the concurrent agreement between the two variables
and gaining insights into the unfolding and variability of of the measured and predicted HPC-CS. In conclusion,
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 14 of 18

12

10

3
1

-M

-M

-M

M
-M

-M

-M

F-

F-

F-
RX

RX

RX
NN

NN

NN

R
A

A
R

N
G

CALIBRATION PHASE MAPE

VERIFICATION PHASE MAPE
Fig. 6 Error plot for the standalone models in the first scenario

Table 6 Result of the Ensemble Algorithm

Models Calibration Phase Verification Phase
R PCC MSE MAPE MAE R PCC MSE MAPE MAE

RBFNN-NARX 0.955 0.955 0.001 0.992 0.006 0.743 0.744 0.002 2.144 0.010
RBFNN-GRNN 0.862 0.871 0.002 4.432 0.014 0.455 0.593 0.004 3.998 0.018
RBFNN-RF 0.961 0.961 0.001 0.952 0.006 0.968 0.972 0.000 0.987 0.005

ensemble RBFNNs are adaptive, allowing for more flex- box-and-whisker plot, which is a visual depiction of the
ible and adaptable solutions to difficult problems like CS. distribution of a dataset. It provides a concise overview
RBFNN’s success in concrete prediction, particularly for of the data’s central tendency, spread, and asymmetry.
HPC-CS, is a result of its superiority to GRNN, NARX- The box within the plot signifies the interquartile range
NN and RF, as indicated in Table 6. The original scenar- (IQR), encapsulating the middle 50% of the data. The line
io’s numerical and visual demonstrations of the recital inside the box denotes the median, or the central value
proficiencies of various paradigms show that the models of the dataset. The whiskers extend from the box’s edges
failed to simulate HPC-CS in both model combinations. to reveal the data’s range, excluding any outliers, which,
Different ensembled paradigms were then designed to if present, are plotted as individual points beyond the
capture and recreate the complex behaviour of HPC-CS whiskers. Box plots are valuable for comparing distribu-
in the first, second and third model combinations (C1, C2 tions, identifying outliers, and gaining insights into the
and C3) by mixing the linear and non-linear behaviour of dispersion and variability of datasets, making them a
various paradigms. potent tool in statistical analysis and data visualization.
According to their unique MAPE values, various mod-
els are graphically compared in Fig. 9. As observed, the 5 Conclusion
performance of the models would degrade with increas- The most important mechanical property of concrete is
ing error counts, and vice versa. As can be seen in Fig. 9 its compressive strength (CS). Therefore, poor CS can
for the graphical performance, RF delivers higher per- result in severe failures and is exceedingly challeng-
formance accuracy than GRNN and NARX-NN. MAPE ing to fix. Thus, proper and early prediction of concrete
measurements are described below in further detail. strength is a significant task for academics and concrete
As depicted in Fig. 9, the double-error bar chart on designers. Due to tremendous complexity of the High-
the vertical axis represents a box plot, also known as a Performance Concrete (HPC), simulating its behaviour
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 15 of 18

Fig. 7 Scatter plot

is difficult. Conventional strategies for proportioning in • Combination M1 of the RF model exhibited supe-
High-performance concrete (HPC) are confined to their rior performance in the first scenario, delivering
effectiveness, because of cost, utilization constraints, and high prediction accuracy, with PCC = 0.854 and
nonlinear relationships. This examination is aimed to MAPE = 4.349 during the calibration phases, as well
address these limitations by implementing a strategy to as PCC = 0.643 and MAPE = 3.008 in the verification
optimize mixtures, thereby reducing layout prices, saving phase.
time, and diminishing fabric wastage in the production • Furthermore, the ensemble of the RF model out-
sector. performed all other models in the second scenario,
By employing various machine learning models, with PCC = 0.961 and MAPE = 0.952 through-
including the Generalized Regression Neural Network out the calibration phase, as well as PCC = 0.972
(GRNN), Nonlinear AutoRegressive with eXogenous and MAPE = 0.987 in the verification phase.
inputs (NARX neural network), Random Forest (RF), and • Overall, the machine learning models proposed in
Radial Basis Function Neural Network (RBFNN), this this study demonstrate their effectiveness as valuable
study estimated the Compressive Strength (CS) of HPC. tools for predicting the CS of HPC. These findings
Based on the modelling results, the following conclusion suggest that these models can significantly contrib-
can be drawn: ute to improving the efficiency and accuracy of HPC
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 16 of 18

Actual CS Predicted RBFNN-GRNN

1.0 .8

.7
0.8
.6
0.6 .5

0.4 .4

.3
0.2
.2

0.0 .1
10 20 30 40 50 60 70 10 20 30 40 50 60 70

Predicted RBFNN-NARX Predicted RBFNN-RF

1.0 1.2

0.8 1.0

0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0 0.0

-0.2 -0.2
10 20 30 40 50 60 70 10 20 30 40 50 60 70
Fig. 8 Time-series plot

4.5

4.0

3.5

3.0

2.5

2.0

1.5

1.0

0.5
RBFNN-NARX RBFNN-GRNN RBFNN-RF

CALIBRATION PHASE MAPE

VERIFICATION PHASE MAPE
Fig. 9 Error plot for the ensemble models
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 17 of 18

mixture optimization, so as to enhance the processes Al-Shamiri, A. K., Kim, J. H., Yuan, T. F., & Yoon, Y. S. (2019). Modeling the
compressive strength of high-strength concrete: an extreme learning
and outcomes of construction. approach. Const Build Mater, 208, 204–219. https://​doi.​org/​10.​1016/j.​
conbu​ildmat.​2019.​02.​165
Behnood, A., Behnood, V., Modiri Gharehveran, M., & Alyamac, K. E. (2017).
Acknowledgements Prediction of the compressive strength of normal and high-performance
Humans and Machines LTD and Civil Lake LTD are given credit by the authors concretes using M5P model tree algorithm. Const Build Mater, 142,
for this research. 199–207. https://​doi.​org/​10.​1016/j.​conbu​ildmat.​2017.​03.​061
Boussaada, Z., Curea, O., Remaci, A., Camblong, H., & Bellaaj, N. M. (2018). A
Author contributions nonlinear autoregressive exogenous (NARX) neural network model for
Each author contributed equally to this study. the prediction of the daily direct solar radiation. Energies. https://​doi.​org/​
10.​3390/​en110​30620
Funding Bui, D., Nguyen, T., Chou, J., Nguyen-xuan, H., & Duc, T. (2018). A modified firefly
Not offered. algorithm-artificial neural network expert system for predicting compres-
sive and tensile strength of high-performance concrete. Const Build Mater,
Availability of data and materials 180, 320–333. https://​doi.​org/​10.​1016/j.​conbu​ildmat.​2018.​05.​201
On-demand access to all information and resources is available. Cheng, M. Y., Firdausi, P. M., & Prayogo, D. (2014). High-performance concrete
compressive strength prediction using genetic weighted pyramid opera-
Code availability tion tree (GWPOT). Engineering Applications of Artificial Intelligence, 29,
Accessible upon demand. 104–113. https://​doi.​org/​10.​1016/j.​engap​pai.​2013.​11.​014
Doddy, P., Tjong, W. F., & Daniel, T. (2018). Prediction of high-performance
concrete strength using a hybrid artificial prediction of high-performance
Declarations concrete strength using a hybrid artificial intelligence approach. MATEC
Web of Conf. https://​doi.​org/​10.​1051/​matec​conf/​20182​03060​06
Ethics approval and consent to participate
Elkiran, G., Nourani, V., & Abba, S. I. (2019). Multi-step ahead modelling of river
Not.
water quality parameters using ensemble artificial intelligence-based
approach. Journal of Hydrology, 577(July), 123962. https://​doi.​org/​10.​
Consent for publication
1016/j.​jhydr​ol.​2019.​123962
Certainly.
Feng, Y., Mohammadi, M., Wang, L., Rashidi, M., & Mehrabi, P. (2021). Application
of artificial intelligence to evaluate the fresh properties of self-consolidat-
Competing interests
ing concrete. Materials. https://​doi.​org/​10.​3390/​ma141​74885
The authors assert that there are no interests at odds.
Gholamrezaei, M., & Ghorbanian, K. (2007). Rotated general regression neural
network. IEEE Int Conf Neural Networks—Conf Proc, 2(6), 1959–1964.
Author details
1 https://​doi.​org/​10.​1109/​IJCNN.​2007.​43712​58
Faculty of Engineering, Department of Civil Engineering, Bayero University
Han, Q., Gui, C., Xu, J., & Lacidogna, G. (2019). A generalized method to predict
Kano, Kano, Nigeria. 2 School of Energy, Geoscience, Infrastructure and Society,
the compressive strength of high-performance concrete by improved
Institute for Sustainable Built Environment, Heriot-Watt University, Edinburgh,
random forest algorithm. Const Build Mater, 226, 734–742. https://​doi.​org/​
UK. 3 Faculty of Science Education, Department of Physics Education, Kano
10.​1016/j.​conbu​ildmat.​2019.​07.​315
University of Science and Technology, KUST, Wudil, Kano, Nigeria. 4 Depart-
Hannan, A., Hannan, S. A., & Manza, R. R. (2010). Generalized regression neural
ment of Civil Engineering, Federal University, Dutsin‑Ma, Katsina, Nigeria.
5 network and radial basis function for heart disease diagnosis generalized
Operational Research Center in Healthcare, Near East University, Nicosia,
regression neural network and radial basis function for heart disease
Turkish Republic of Northern Cyprus. 6 Department of Analytical Chemistry,
diagnosis. Int J Comput Appli. https://​doi.​org/​10.​5120/​1325-​1799
Faculty of Pharmacy, Near East University, Nicosia, Turkish Republic of North-
Jibril, M. M., Bello, A., Aminu, I. I., Ibrahim, A. S., Bashir, A., Malami, S. I., Habibu,
ern Cyprus. 7 Faculty of Engineering, Department of Civil Engineering, Kano
M. A., & Magaji, M. M. (2022). An overview of streamflow prediction using
University of Science and Technology, KUST, Wudil, Kano, Nigeria. 8 Depart-
random forest algorithm. GSC Adv Res Rev. https://​doi.​org/​10.​30574/​
ment of Civil Engineering, Prince Mohammad Bin Fahd University, 31952 Al
gscarr.​2022.​13.1.​0112
Khobar, Saudi Arabia.
Jibril, M. M., Idris, S., Hauwa, M., Jibrin, B., Jibrin, U., Adamu, M., Usman, D. A. G.,
Dilber, A. G. I., Ozsahin, U., & Karami, Z. (2023d). New random intelligent
Received: 24 October 2023 Revised: 11 October 2024 Accepted: 19
chemometric techniques for sustainable geopolymer concrete : low—
October 2024
energy and carbon - footprint initiatives. Asian J Civil Eng. https://​doi.​org/​
10.​1007/​s42107-​023-​00908-7
Jibril, M. M., Malami, S. I., Jibrin, H. B., Muhammad, U. J., Duhu, M. A., Usman,
A. G., Ibrahim, A. G., Ozsahin, D. U., Lawal, Z. K., & Abba, S. I. (2023c). New
random intelligent chemometric techniques for sustainable geopolymer
References concrete: Low-energy and carbon-footprint initiatives. Asian J Civil Eng.
Abba, S. I., Pham, Q. B., Saini, G., Linh, N. T. T., Ahmed, A. N., Mohajane, M., https://​doi.​org/​10.​1007/​s42107-​023-​00908-7
Khaledian, M., Abdulkadir, R. A., & Bach, Q. V. (2020). Implementation of Jibril, M. M., Malami, S. I., Muhammad, U. J., Bashir, A., Usman, A. G., Salami, B.
data intelligence models coupled with ensemble machine learning for A., Rotimi, A., Ibrahim, A. G., & Abba, S. I. (2023a). High strength concrete
prediction of water quality index. Environmental Science and Pollution compressive strength prediction using an evolutionary computa-
Research. https://​doi.​org/​10.​1007/​s11356-​020-​09689-x tional intelligence algorithm. Asian J Civil Eng. https://​doi.​org/​10.​1007/​
Abisek, S., Raja, A., Muthusamy, S., & Panchal, H. (2022). Prediction of strength s42107-​023-​00746-7
and analysis in self-compacting concrete using machine learning based Jibril, M. M., Zayyan, M. A., Idris, S., Usman, A. G., Salami, B. A., Rotimi, A., & Abba,
regression techniques advances in engineering software prediction S. I. (2023b). Applications in Engineering Science Implementation of non-
of strength and analysis in self-compacting concrete using machine linear computing models and classical regression for predicting compres-
learning based regression tec. Adv Eng Software, 173(November), 103267. sive strength of high-performance concrete. Appl Eng Sci. https://​doi.​org/​
https://​doi.​org/​10.​1016/j.​adven​gsoft.​2022.​103267 10.​1016/j.​apples.​2023.​100133
Aliyu, D. S., Malami, S. I., Anwar, F. H., Farouk, M. M., Labbo, M. S., & Abba, S. I. Li, H., Guo, S., Li, C., & Sun, J. (2013). Knowledge-based systems a hybrid annual
(2021). Prediction of compressive strength of lightweight concrete made power load forecasting model based on generalized regression neural
with partially replaced cement by animal bone ash using artificial neural network with fruit fly optimization algorithm. Knowledge-Based Syst, 37,
network. Int Conf Multidisc Eng Appl Sci. https://​doi.​org/​10.​1109/​ICMEA​ 378–387. https://​doi.​org/​10.​1016/j.​knosys.​2012.​08.​015
S52683.​2021.​96923​17
Muhammad et al. AI in Civil Engineering (2024) 3:21 Page 18 of 18

Lim, C. H., Yoon, Y. S., & Kim, J. H. (2004). Genetic algorithm in mix proportion-
ing of high-performance concrete. Cement Concrete Res, 34(3), 409–420.
https://​doi.​org/​10.​1016/j.​cemco​nres.​2003.​08.​018
Liu, Y. (2022). High-Performance Concrete Strength Prediction Based on. 2022.
Malami, S. I., Akpinar, P., & Lawan, M. M. (2018). Preliminary investigation of car-
bonation problem progress in concrete buildings of north Cyprus. MATEC
Web of Conf, 203, 1–9. https://​doi.​org/​10.​1051/​matec​conf/​20182​03060​07
Mayet, A. M., Al-Qahtani, A. A., Qaisi, R. M. A., Ahmad, I., Alhashim, H. H., &
Eftekhari-Zadeh, E. (2022). Developing a model based on the radial basis
function to predict the compressive strength of concrete containing Fly
Ash. Buildings. https://​doi.​org/​10.​3390/​build​ings1​21017​43
Moodi, Y., Ghasemi, M., & Mousavi, S. R. (2022). Estimating the compressive
strength of rectangular fiber reinforced polymer–confined columns using
multilayer perceptron, radial basis function, and support vector regres-
sion methods. J Reinforc Plast Compos, 41(3–4), 130–146. https://​doi.​org/​
10.​1177/​07316​84421​10501​68
Mustapha, R., & Mohamed, E. A. (2017). High-performance concrete compres-
sive strength prediction based weighted support vector machines. Inter-
national Journal of Engineering Research and Applications, 07(01), 68–75.
https://​doi.​org/​10.​9790/​9622-​07010​16875
Nikoo, M., Torabian Moghadam, F., & Sadowski, Ł. (2015). Prediction of concrete
compressive strength by evolutionary artificial neural networks. Advances
in Materials Science and Engineering. https://​doi.​org/​10.​1155/​2015/​849126
Nourani, V., Elkiran, G., & Abba, S. I. (2018). Wastewater treatment plant per-
formance analysis using artificial intelligence—an ensemble approach.
Water Science and Technology, 78(10), 2064–2076. https://​doi.​org/​10.​2166/​
wst.​2018.​477
Ouyang, H. T. (2017). Nonlinear autoregressive neural networks with external
inputs for forecasting of typhoon inundation level. Environmental Moni-
toring and Assessment. https://​doi.​org/​10.​1007/​s10661-​017-​6100-6
Page, J. W., Cadasse D., Fontana A. (1988). Advanced in Engineering and
Intelligence Systems Estimation of fresh and hardened properties of self-
compacting concrete by optimized radial basis function methods.
Pazouki, G., Golafshani, E. M., & Behnood, A. (2022). Predicting the compressive
strength of self-compacting concrete containing Class F fly ash using
metaheuristic radial basis function neural network. Struct Concrete, 23(2),
1191–1213. https://​doi.​org/​10.​1002/​suco.​20200​0047
Pengcheng, L., Xianguo, W., Hongyu, C., & Tiemei, Z. (2020). Prediction of
compressive strength of high-performance concrete by random forest
algorithm. IOP Conf Series Earth Environ Sci. https://​doi.​org/​10.​1088/​1755-​
1315/​552/1/​012020
Sufian, M., Qian, Y., Su, M., & Hakamy, A. (2023). various parameters on the
flexural strength of ultra-high-performance application of machine learn-
ing algorithms to evaluate the in fl uence of various parameters on the fl
exural strength of ultra-high-performance concrete. Front Mater. https://​
doi.​org/​10.​3389/​fmats.​2022.​11145​10
Vidivelli, B., & Jayaranjini, A. (2016). Prediction of compressive strength of high
performance concrete containing industrial by products using artificial
neural networks. Int J Civil Eng Technol, 7(2), 302–314.
Yu, Y., Li, W., Li, J., & Nguyen, T. N. (2018). A novel optimised self-learning
method for compressive strength prediction of high performance
concrete. Const Buil Mater, 184, 229–247. https://​doi.​org/​10.​1016/j.​conbu​
ildmat.​2018.​06.​219
Zimit, A. Y., Jibril, M. M., Azimi, M. S., & Abba, S. I. (2023). Journal of the saudi
society of agricultural sciences hybrid predictive based control of
precipitation in a water-scarce region : a focus on the application of intel-
ligent learning for green irrigation in agriculture sector. J Saudi Soc Agricu
Scinces. https://​doi.​org/​10.​1016/j.​jssas.​2023.​06.​001

Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in pub-
lished maps and institutional affiliations.