The Efficiency of Using Machine Learning Techniques in Fiber-Reinforced-Polymer Applications in Structural Engineering

Abstract

Sustainable solutions in the building construction industry have emerged as a new method for retrofitting applications in the last two decades. Fiber-reinforced polymers (FRPs) have garnered much attention among researchers for improving reinforced concrete (RC) structures. The existing design guidelines for FRP-strengthened RC members were developed using empirical methods that are based on specific databases, limiting the accuracy of the predicted results. Therefore, the use of innovative and efficient prediction tools to predict the behavior of FRP-strengthened RC members has become essential. During the last few years, efforts have been progressively focused on the use of machine learning (ML) as a feasible and effective technique for solving various structural engineering problems. Its capability to predict the behavior of complex nonlinear structural systems while considering a wide range of parameters offers a distinctive opportunity to make the behavior of RC members more predictable and accurate. This paper aims to evaluate the current state of using various ML algorithms in RC members strengthened with FRP to enable researchers to determine the capabilities of current solutions as well as to find research gaps to carry out more research to bridge revealed knowledge and practice gaps. Scopus databases were searched using predefined standards. The search revealed ninety-six articles published between 2016 and 2023. Consequently, these articles were analyzed for ML applications in the field of FRP retrofitting, including flexural and shear strengthening of RC beams, flexural strengthening of slabs, confinement and compressive strength of columns, and FRP bond strength. The results reveal that 32% of the reviewed studies focused on the application of ML techniques to the flexural and shear strengthening of RC beams, 32% on the confinement and compressive strength of columns, 6.5% on the flexural strengthening of slabs, 22% on FRP bond strength, 6.5% on materials, and 1% on beam–column joints. This research also revealed that the application of various ML algorithms has shown a significant improvement in resistance prediction accuracy as compared with the existing empirical solutions. Supervised learning techniques were the most favorable learning method due to their good generalization, interpretability, adaptability, and predictive efficiency. In addition, the selection of suitable ML algorithms and optimization techniques is found to be mainly dictated by the nature of the problem and the characteristics of the dataset. Nonetheless, selecting the most appropriate ML model and optimization algorithm for each specific application remains a challenge, given that each algorithm is developed with different principles and methodologies.

1. Introduction

A combination of steel and concrete materials forms the reinforced concrete (RC) structures that are being widely adopted by buildings, bridges, and marine engineering. The use of steel in RC structures has significant issues related to corrosion [1], which could lead to performance degradation [2]. Furthermore, the use of composite materials in reinforced concrete structures has become popular among the research community due to the drawbacks of conventional techniques (concrete and steel jacketing) such as corrosion, durability, time consumption, and complex applications. Fiber-reinforced polymer (FRP) jackets have garnered much attention among all jacketing techniques due to their advantageous properties (i.e., corrosion resistance, a high strength-to-weight ratio, simplicity of application and speed, and minimal geometrical change) [3,4].

The most common types of FRP bars are carbon fiber-reinforced polymer (CFRP), glass fiber-reinforced polymer (GFRP), basalt fiber-reinforced polymer (BFRP), aramid fiber-reinforced polymer (AFRP), steel–FRP composite bars (SFCBs), and hybrid fiber-reinforced polymer (HFRP). Despite the general utilization of CFRPs for critical structural component reinforcement and the high cost of carbon fibers, CFRPs have excellent mechanical properties compared with steel [5,6]. GFRP bars have low production costs due to their vast source of raw materials and mature production technology, which has led to their wide use in civil engineering. BFRP bars have similar mechanical properties to GFRP and are considered environmentally friendly as basalt fibers are produced from melted basalt rocks [7,8]. The combination of basalt or glass fibers and carbon fibers forms HFRPs, which are ideal for applications involving humidity and high temperatures [9]. AFRP bars are rarely used in civil engineering practice due to their poor durability and hydrophilic nature.

The existing design guidelines for FRP-strengthened RC members were developed using empirical methods that are based on specific databases, limiting the accuracy of the predicted results. Therefore, the use of innovative and efficient prediction tools to predict the performance of FRP-strengthened RC members has become essential. ML, as a subset of artificial intelligence (AI), aims to teach computers to predict outcomes from available algorithms and datasets. Despite being introduced in 1943, machine learning (ML) only started to thrive in the 1990s. The applications of ML in many real-world scenarios, such as medical diagnosis, traffic alerts, speech and image recognition, and self-driving cars [10,11,12,13] have led ML to be the most successful area of AI and one of the technology buzzwords of our age.

The shortcoming of existing design guidelines for FRP-strengthened RC members and the increase in adoption of ML techniques have led to the use of a significant number of ML applications to predict the behavior of FRP-strengthened RC members, e.g., neural networks (NNs), linear regression (LR), support vector machine (SVM), and standard ensemble learning (EL) models including random forest (RF) and extreme gradient boosting (XGBoost). Therefore, it is noted that the application of ML in the field of FRP strengthening of RC structures is currently in a constant state of development and refinement while further considering structural and environmental factors [14].

Very few attempts have been made to analyze ML applications to predict the performance of specific RC members strengthened with FRP and have either focused on a single RC member or different composite materials [15,16]. For example, Sandeep et al. [15] provided a detailed review on the application of ML for predicting the shear strength of beams reinforced with FRP bars, whereas the ML applications for the structural behavior of members made of various materials (e.g., steel, concrete, cold-form steel (CFS), steel–concrete composite, and FRP composite) were thoroughly evaluated [16]. However, the current state of research still lacks a detailed discussion on the implementation of ML to critically evaluate the efficiency of FRP in strengthening various RC structural members. As a result, there is a need for such a study to critically analyze the existing and future applications of ML to enhance the efficiency of employing FRP for strengthening various RC structural members.

This paper presents a comprehensive state-of-the-art review of the application of ML in the field of FRP strengthening of RC structural members, with a focus on the efficacy of such a technique as compared with existing empirical methods. Within this scope, the current study explores a variety of ML-FRP applications, such as shear and flexural strengthening of RC beams, confinement of columns, strengthening of slabs, beam–column joints, and FRP–concrete interfacial strength. In light of the aforementioned focus, this review discusses the findings of previous research, the relevance of the approaches used, the limitations, and future directions. This will allow the research community to identify knowledge gaps in significant published articles as well as the maturity level of the present approaches and to then strive to either bridge the gaps or improve the maturity level of existing applications.

To fill the research gap and to achieve the research objectives, firstly, published articles were collected using the relative keywords and a Scopus database (Section 2). Secondly, a scientometric analysis was used to look at the relationships between published articles (topics) to highlight the most advanced areas of applications (Section 3); then, a thematic analysis was used to categorize the published articles into particular themes; and a gap analysis was performed to examine the most important published articles in each field in terms of their methodology, findings, and limitation(s) for each application (Section 4). This was followed by highlighting the crucial implemented parameters (Section 5). Finally, the results are discussed in Section 6, and subsequently, the conclusions and future works are drawn in Section 7.

2. Methodology and Logic

The research methodology, including the data collection and analysis, i.e., scientometric, thematic, and gap analysis research, is shown in Figure 1. A Scopus database was utilized to search for relevant articles. Given that, this paper discusses the applications of ML in the FRP strengthening of reinforced concrete members. Consequently, specific keywords were used to find relevant articles, such as (TITLE-ABS-KEY (“fiber reinforced concrete”) OR TITLE-ABS-KEY (“FRP”) AND TITLE-ABS-KEY (“machine learning”) AND TITLE-ABS-KEY (“reinforced concrete”) AND (LIMIT-TO (PUBSTAGE, “final”) AND (LIMIT-TO (DOCTYPE, “ar”)).

Following the above search for relevant articles, a scientometric analysis was first utilized, as it is considered an effective approach to gauge the advancement of scientific production and to define the overlapping interests of bibliometrics and informatics [17]. Furthermore, a scientometric analysis was performed to examine the connections between these various application areas and the density of each application. Subsequently, a gap analysis was implemented, as it can be used to detect missing elements in any study, literature review, or program analysis [18]. To conduct the gap analysis, themes were identified through techniques such as scanning and skimming to sort the relevant articles [19]. As a consequence, the findings were evaluated and divided into major themes and sub-themes for critical analysis in order to emphasize the purpose, approach, and limitations of the study.

3. Scientometric Analysis

This paper has deployed a quantitative-based science mapping technique. According to Cobo et al. (2011) [20], this could help in quantitatively mapping out patterns and networks in a large set of bibliometric data. Therefore, “science mapping” was the first primary adopted method. This method was implemented because of its capability to picture large bibliographical units and systematic patterns in bodies of literature. A bibliometric analysis aims to systematically chart the literature per se; scientometric scanning extends the bibliometric analysis to cover the analysis and measurement of the researchers, institutions, and countries within the literature [21]. Therefore, a scientometric analysis uses bibliometric data, techniques, and methods to scientifically map out the literature [22]. The progress of scientific production in the applications of ML in FRP strengthening of RC members is shown in Figure 2. The use of ML approaches in this field has been demonstrated to have begun in 2016 and has been significantly rising since 2020. This shows how much attention ML applications with FRP have garnered during the past four years.

The majority of selected articles are published in top-ranked journals, as shown in

Table 1. An example of reliable sources used to collect data.

Journal Name	Number of Publications
Engineering Structures	9
Composite Structures	9
Polymers	6
Construction and Building Materials	5
Materials	5
Structures	5

; for example, nine articles were published in the Engineering Structures journal, nine articles were published in the Composite Structures journal, and another six articles were published in Polymers.

Figure 3 illustrates the network analysis of n = 91 papers. Four clusters were identified and connected, and it can be seen from Figure 3 that there are specific usages of integrating ML and FRP in strengthening reinforced concrete, namely, optimizing the design of isolated structural members through the prediction of different parameters, i.e., flexural strength, shear strength, tensile strength, bond strength, and punching shear. The density of applications described in the analysis of 91 articles is shown in Figure 4. It is clear that ML technology is heavily integrated with FRP applications to provide a wide variety of practical solutions.

4. Results

This section’s hierarchy represents challenges focusing on the use of ML to analyze and improve RC members utilizing FRP. The main themes revealed from the review were FRP-strengthening concrete structure (beams, columns, slabs, and beam–column joints), FRP–concrete interface (bond strength), and FRP materials. The results revealed that 32% of the reviewed studies focused on the application of ML techniques on flexural and shear strengthening of RC beams, 32% focused on confinement and compressive strength of columns, 6.5% focused on flexural strengthening of slabs, 22% focused on FRP–concrete bond strength, 6.5% focused on materials, and 1% focused on beam–column joints. Figure 5 represents the main theme and sub-theme hierarchy of the results, including the percentage of each of the themes. All ML models discussed in the following sections are listed in Appendix A.

4.1. Beams

Reinforced concrete beams are an essential type of structural element that plays a key role in how weight are transmitted and guarantee that a building’s foundation is securely in the ground. The results reveal that 32% of the reviewed studies focused on the application of ML techniques to the flexural and shear strengthening of RC beams. This section discusses various ML algorithms used in the literature for predicting the shear and flexural capacities of RC beams.

4.1.1. Shear Strength

The shear capacity of RC beams was predicted mathematically using a variety of ML approaches [23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42]. The use of the well-known artificial neural network (ANN) technique was adopted to investigate the impact of various crucial variables on the shear strength of FRP-RC beams [25]. Two ANN models were developed, one involving all independent parameters and the other based on specific pre-determined variables. It was found that the model with selected parameters resulted in substantially higher accuracy as compared with the model considering all variables. This indicates that input variables used in the development of the ML models have a significant effect on the model’s performance and, consequently, the prediction accuracy and concluded outcomes. In a similar study, Abuodeh et al. [23] performed a parametric study to determine the most influential parameters on the shear capacity of FRP-strengthened RC beams. The resilient back-propagating neural network (RBPNN) algorithm was employed to analyze 120 tested beams from the literature. Based on the parametric study, the RBPNN yielded relatively accurate results that were heavily dependent on the variables selected. Furthermore, two expressions to estimate the shear strength of RC beams strengthened with FRP sheets were established using an ANN model fused with the GA algorithm [33,34]. The accuracy of the developed expressions was validated using extensive datasets, showing the reliability of the approach used. The ANN algorithm was also adopted to estimate the FRP shear contribution in RC beams [35]. Using a dataset of 511 tested beams collected from the literature, the results revealed that ANN yielded reasonable and accurate results, indicating the feasibility of the ML model.

The use of various ANN algorithms, including ANN-LM, ANN-QN, ANN-CG, and ANN-GD, to develop accurate and more robust models to predict the shear strength of FRP-RC beams without stirrups was recently investigated [36]. Based on 307 tested beams from the literature, it was shown that ANN can accurately predict the shear strength of RC beams with a correlation coefficient (R) of 0.961. In addition, Marani and Nehdi [28] developed a shear model using the tabular generative adversarial network (TGAN) algorithm. Based on an extensive dataset of 304 experiments, the results demonstrated that the TGAN technique yielded superior accuracy in comparison with existing design codes, with an R² of 0.96. Most recently, an ANN model was developed for the shear strength prediction of RC beams strengthened with EB-FRP jackets [37,38]. Based on extensive datasets collected covering a wide range of parameters, the ANN model had high prediction accuracy, confirming the potential of employing the ML technique in the development of current design codes.

The application of the extreme learning machine (ELM) method was also implemented to predict the ultimate shear resistance of RC beams retrofitted with FRP rebars [39]. It was shown that the ELM can be utilized to accurately measure the shear strength of composite-strengthened RC beams. In a more recent study [32], the predictive efficacy of three ML methods, namely, ELM, M5-Tree, and RF, was compared using the results on 112 FRP-strengthened RC beams. The results revealed that the M5-Tree model yielded slightly higher accuracy in predicting the shear strength as compared with the other two models.

Using ensemble approaches, the shear strength of FRP-enhanced RC beams was investigated as well. Kaveh et al. [26] employed a decision-tree-based algorithm that involves XGBoost and compared the results with the current design codes and other supervised ML techniques, namely, LASSO and RF. Based on 205 RC beams strengthened with FRP with no internal stirrups, the XGBoost model was superior in terms of shear strength prediction with higher generalized performance. More recently, Wakjira et al. [30] conducted a comprehensive study on the shear strength prediction of FRP-RC beams. In this study, various single and ensemble ML models were implemented. It was reported that all ensemble models showed superior prediction capabilities and resulted in significantly higher prediction accuracy than the single model. More specifically, the XGBoost model outperformed all single and ensemble models as well as the current design guidelines in estimating the shear capacity of the beams. The use of XGBoost was also implemented to predict the shear capacity of FRP-RC beams with and without stirrups [27]. Based on a large dataset encompassing 453 tested specimens collected from the literature, XGBoost yielded remarkable results, with R² values higher than 0.95 confirming the reliability of the model. It was also concluded that XGBoost’s performance, however, depends heavily on the dataset used for the model’s training and testing, which could potentially limit the use of the model when unstructured data are used.

Lately, Rahman et al. [40] generated 10 ML models involving different ensemble techniques. With the use of an extensive database of rectangular and T-beams, the ensemble models RF, CatBoost, and XGBoost showed superior prediction performance in predicting the shear capacity when compared with existing design guidelines. Additionally, Yang and Liu [41] conducted a comparative study of various ML models and existing codified expressions. Several ensemble models were implemented, comprising LR, decision trees (DT), RF, and XGBoost, to estimate the shear resistance of FRP-strengthened RC beams with no internal shear reinforcement. The existing design codes provided conservative results, while the ML-based models yielded reasonably accurate results, particularly when crucial parameters were selected in the development process.

Multiple predictive models were also established using the Gene Expression Programming (GEP) method to study the shear strength of strengthened RC beams with FRP [29,31]. The shear resistance of 388 slender and deep beams strengthened with FRP bars was utilized to develop a robust predictive model using GEP [29]. In comparison with existing predictive models, the developed shear expression was superior, particularly in the case of deep beams. Anvari et al. [31] used GEP to build two models that estimate the shear resistance of RC beams strengthened with FRP sheets. The first model is a generalized one that can be applied to beams with and without shear reinforcement, whereas the second model involves two separate expressions for each case. With the use of a large database of 785 experimental results, the developed models showed high prediction accuracy when compared with corresponding formulas from the literature, indicating the significance of using GEP models.

On the other hand, a new hybrid model that combines the Smart Firefly Algorithm (SFA) and least squares support vector regression (LSSVR) was developed to effectively predict the shear strength of FRP-RC beams [24]. To verify the accuracy of the developed model, the results of the hybrid technique were comprehensively compared with ensemble and single ML methods. It was found that the hybrid model had significantly more accurate predictions, suggesting its feasibility to alter the other two ML techniques.

Recently, the punching shear capacity of FRP-RC slender beams was investigated using hybrid ML models [42]. In this study, RF models were optimized by employing various techniques, including the Ant Lion Optimizer (ALO), the Moth Flame Optimizer (MFO), and the Salp Swarm Algorithm (SSA). The results revealed that the developed hybrid model showed a good correlation with experimental results and had higher precision in predicting the punching shear capacity as compared with other traditional techniques.

The developed ML models, in general, outperformed the existing design guidelines in predicting the shear capacity of FRP-RC beams. The prediction accuracy of such models was further enhanced by pre-determining some of the crucial parameters and adopting optimization algorithms like the genetic algorithm (GA). The applied ML algorithms for beam shear strength estimation are illustrated in Figure 6.

4.1.2. Flexural Strength

By using a modern version of AI that is human-explainable, Naser [43] predicted the maximum moment capacity of FRP-strengthened RC beams and the tendency of the FRP to fail under different mechanisms by using explainable artificial intelligence (XAI) and interpretable machine learning (IML). As a result, three techniques were used, namely, keras deep neural networks (KDNNs), light gradient-boosted trees (LGBTs), and ExGBT. The results revealed that while testing the maximum moment capacity and mode of failure of FRP-strengthened beams, ExGBT showed the best performance. The study comes to the conclusion that structural engineers are still hesitant to use AI and ML approaches as their major tools for designing and analyzing structures. This is because of a few concerns with AI/ML’s opacity and the minimal coding and programming instruction that is offered in most curricula.

Aravind et al. [44] identified the beam cracks by utilizing appropriate ML algorithms and failure pattern identification approaches. To achieve this, M30-grade geopolymer and ordinary concrete beams were created utilizing steel bars, BFRP, and GFRP. The failure modes of the beams were divided into three types, namely, flexure, shear, and compression, using six ML algorithms (i.e., SVM, DT, Gaussian NB, SGD, K-neighbor, and Adaboost). ML classifiers were used to calculate the confusion matrix, precision, accuracy, and recall scores. It was discovered that the support vector classifier, out of the six employed, performed the best, recognizing failure patterns with a 100% accuracy rate.

According to the factors controlled by the material and geometric characteristics, Saleh et al. [45] identified the failure mechanism of fiber-strengthened RC beams. In addition, the authors investigated the deformability capacity of FRP-RC beams to determine whether certain design constraints should be set to ensure that minimal deformability is given by design for a wide variety of FRP-RC beam characteristics. Furthermore, these authors employed SVM, extensive assembled experimental datasets, and a verified analytical model to determine the failure mechanism of FRP-RC beams. The findings indicate that the SVM algorithm performed well in classifying the failure mode of FRP-RC beams. The accuracy of the ACI guideline definition was 87%, compared with 97% for the failure mode categorization. Additionally, the SVM algorithm’s classification rule was not greatly impacted by the random fluctuations in parameter values.

A unified model for predicting the moment bearing capacity of RC beams improved with FRP materials was developed by Zhang et al. [46] to address the problems with different models, challenging calculations, and limited accuracy. In order to anticipate the moment capacity, the experimental data for three popular FRP strengthening types—externally bonded, end anchoring, and near-surface mounted FRP—were gathered from the literature. The prediction model was performed using the XGBoost algorithm. It was compared with prediction models built using two excellent ML algorithms: support vector regression (SVR) and ANN. The predictive model has a considerable generalization ability and can be used based on the prediction of the bearing capacity of FRP-strengthened RC beams. Follow-up studies can continuously improve the accuracy of the model by expanding the dataset and further validating it in actual engineering to provide guidelines for the design and application of beams strengthened by FRP.

Zhang et al. [47] investigated the reliability and feasibility of utilizing EL to predict the flexural capacity of RC beams strengthened with FRP. The study was implemented using four different EL algorithms, including RF, gradient boosting decision tree (GBDT), Adaboost, and XGBoost. The results indicated that EL-based models performed much better than empirical models and ML-based models individually. As a result, the suggested EL-based models showed potential for use in engineering applications. Perera et al. [48] created a novel method that combines electro-mechanical impedance (EMI) with multilevel hierarchical ML techniques. Fiber Bragg grating (FBG) temperature and strain sensors were utilized to measure the mechanical behavior of RC beams strengthened with near surface mounted (NSM) FRP under sustained load and different temperatures. The usefulness of the suggested strategy was proved through the examination of the experimental data in a particularly complicated scenario. Despite the positive outcomes, further research should evaluate the suggested strategy in greater detail. By applying it to additional specimens, considering various types of loading, including fatigue loading, and a more limited range of temperature change, the effectiveness of this strategy in comparison with other strengthening techniques, such as externally bonded reinforcement (EBR), should also be examined. Guo et al. [49] proposed an integrated model based on ML methods to measure the moment capacity and ductility of the compression-yielding (CY) T-beam. The proposed model was implemented using the SVM model, which combines different kernel functions that evaluate the effectiveness of CY beams with T sections. The ductility of a CY T-beam was then predicted using gaussian process regression (GPR). Finally, a genetic algorithm (GA) is created to determine optimal CY beam section design alternatives. The results demonstrated that the integrated model produced highly accurate predictions. Further research is required to include the cost of materials in the design consideration. A sensitivity study of the post-modulus ratio is also strongly advised because of the complicated behavior of the CY beam.

Based on the results of an improved super-learner ML model, Wakjira et al. [50] created an efficient prediction tool for a fast, accurate, and intelligent (FAI) prediction of the flexural capacity of FRP-RC beams. The experimental findings on the flexural strength of FRP-RC beams were assembled into a database and randomly divided into 80% training and 20% test sets. The suggested super-learner ML model’s predictive power was compared with that of boosting- and tree-based ML models. The results showed that all ensemble models performed well in predicting the flexural capacity of FRP-RC beams. The major failure mechanisms of FRP-strengthened RC beams under flexure are plate end (PE) debonding and intermediate crack (IC) debonding. Different failure characteristics are displayed by various failure modes. Hu et al. [51] employed six ML algorithms, including the K-nearest neighbor (KNN), RF, DT, LR, back-propagation neural network (BPNN), and SVM, to predict the failure modes of FRP-strengthened RC beams. The results proved that DT and RF are superior to the other methods for forecasting the failure modes of FRP-strengthened RC beams. Future database expansion is required to provide a more precise model that can identify the debonding mechanisms of FRP-strengthened RC beams under flexure.

The developed ML models were frequently more successful in predicting the flexural capacity of FRP-reinforced beams than the labor-intensive and expensive experimental procedures. The prediction accuracy of these algorithms was greatly improved by pre-determining some of the key parameters and utilizing optimizing ML techniques such as ANN. Figure 7 shows the developed ML algorithms for predicting the flexural strength of FRP-strengthened beams.

4.2. Columns

The results revealed that 32% of the reviewed studies focused on the application of ML techniques to the confinement and compressive strength of columns. Axial compressive capacity, compressive strength of FRP-confined concrete, bridge columns (drift ratio), and fire exposure influence in FRP-strengthened members are the main sub-themes.

4.2.1. Axial Compressive Capacity

Bakouregui et al. [52] represented a novel technique for estimating the load-carrying capacity of RC columns strengthened with FRP bars using XGBoost algorithm. The XGBoost model’s effectiveness and accuracy were assessed and compared with the available equations in the literature and existing design codes. The findings revealed that the suggested prediction model was appropriate for predicting the load-carrying capacity of FRP-RC columns. However, the XGBoost model had several constraints such as the completeness of the data, quantity, quality, and distribution of the input parameters. Therefore, these constraints have a significant impact on the predictive model’s performance. The FRP-RC column database was created from 36 different sources from published research. As a result, as new experimental data become available, the database should be updated.

Cakiroglu et al. [53] used the power of AI to build a novel technique for predicting the axial capacity of FRP-RC columns utilizing data-driven ML algorithms. EL approaches, such as GBRF and XGBoost, were shown to accurately estimate axial load-carrying capability. The coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error were used to assess the accuracy of eight different ML models’ predictions. The SHAP method was used to determine the factors that have shown the greatest influence on structural response based on the prediction model with the highest performance. The proposed equations were established according to an experimental database of 117 samples. Furthermore, the outcomes expected by the derived equations were valid within the database’s range. The suggested equations were data-driven; therefore, the accuracy of mechanics-based capacity prediction model is based on the data accuracy. It is suggested that in combination with experimental research, the databases might be improved with the use of highly accurate finite element models. Future research in this field should be focused on axial load-carrying capacity prediction under eccentric axial loading and database expansion for model training.

Ma et al. [54] predicted the axial compressive capacity of concrete-filled steel tubular (CFST) short columns confined with CFRP utilizing an innovative ML algorithm, XGBoost. Eight algorithms—LR; KNN; SVM; and typical EL models such as RF, adaBoost, GBDT, XGBoost, and lightGBM regressor (LGB)—were selected for the calculations and compared. The XGBoost model was optimized because it had excellent prediction ability for the axial compressive capacity of strengthened CFST columns with CFRP. Peak loads of the CFST specimens’ actual and expected values were compared, and the results showed that the suggested XGBoost model had outstanding predictive capabilities.

Tarawneh et al. [55] proposed an ANN-based model capable of estimating the axial capacity and slenderness limit as well as developing an interaction diagram for FRP-reinforced columns. The previously described model was trained using Bayesian regularization on a wide database of 241 tested columns. The predictions of the ANN-based model, which had a COV of 15% and a root-mean-square error of 130 kN, closely matched the experimental results of the generated database.

Arora et al. [56] investigated the axial load-carrying capacity (ALCC) of FRP-reinforced concrete columns. An innovative and trustworthy ANN model based on ML was developed to efficiently predict the ALCC of FRP-reinforced concrete columns. The types of concrete investigated in this study include normal-weight concrete and geopolymer-based concrete. The created model was efficient and simple to use for determining the axial capacity of FRP-reinforced concrete columns. However, the created ANN model had the limitation of only being applicable to values within the input and output ranges. Because the current model was constructed using only 242 experimental specimens, more experimental data should be added, and the model’s quality should be enhanced to expand its precision and allowable range of input and output parameters. Moreover, some of the less important input parameters could be deleted from the database to boost the model’s dependability even more.

An ML model was created by Miao et al. [57] to estimate the ultimate strength of circular concrete-filled FRP–steel composite tube (CFSCT) columns. A comprehensive dataset of 305 samples of circular CFSCT columns under axial stresses from published literature was used to train and evaluate SVR, BPNN, and RF machine learning techniques. As a consequence of their improved prediction accuracy and applicability compared with the current empirical models, the results suggested that ML-based models were an effective alternative for the current empirical solutions. However, due to the RF model’s over fitting issue, only the SVR and BPNN algorithms were regarded as being appropriate for the final load prediction. The ML-based model created for this study had limitations due to the small dataset. As datasets become more complicated, advanced ML algorithms could be more useful and precise. For members who do not have access to experimental data, FE models can be used to create data samples. The most practical and cost-effective approach to overcome the limitations of small datasets is through combining FE and ML-based models.

A novel model was built by Almomani et al. [58], utilizing a promising GEP version to predict the axial capacity of FRP-RC columns. The behavior of concentrically short FRP-RC columns has been extensively studied in recent years; however, little attention has been paid to the influence of load eccentricity and the slenderness ratio of FRP-RC columns. Furthermore, there is still disagreement on the approaches used to account for the impact of column slenderness because no conclusive evidence was provided. Two GEP models were developed for estimating the axial capacity based on load eccentricity using experimental data of FRP-RC columns that were collected from the literature. The results showed that in terms of anticipating axial capacity, the suggested models demonstrated great accuracy.

Shin and Park [59] established a hybrid optimization technique (GA-ANN) that combines genetic algorithms (GAs) and ANN to quickly determine the best retrofit schemes for RC columns in frame buildings without the time-consuming nature of manually repeated operations. Rapid data production advantages of ANN and optimization solution advantages of GA were coupled. This was to enhance and strengthen non-ductile RC frame structures under multi-hazard loads (such as blast and seismic loads). Therefore, the suggested quick decision-making technique was utilized to instantly select the best FRP retrofitting strategy. The following retrofit characteristics were investigated: the FRP jacket thickness and strength, grout strength, and inner diameter of columns. The machine-learning-based technology maximized the confinement ratio and minimized the stiffness ratio to optimize the retrofit details within target performance levels, and it predicted allowable seismic load ranges. Based on the analysis, stiffness characteristics relating to geometric conditions at low confinement levels were improved rather than confinement parameters.

Recycled aggregate concrete (RAC), despite its many advantages, typically has inferior characteristics to natural aggregate concrete, which was seen as a challenge to its wide availability and usage. FRP jacketing is one method to overcome this problem. However, there is difficulty in assessing the axial performance of FRP-confined recycled aggregate concrete columns (FRACCs), partly because of the complicated load-resisting systems involved. The matter is further complicated by RAC’s inherent weakness. A different approach to resolving this issue was through Zhao et al.’s work [60]. In order to expand the experimental database, TGAN, a synthetic data generator, was used. After that, the settings of the XGBoost model were created by utilizing the BAS algorithm. The created model performed better than a number of standard ML models as well as several of the current empirical equations. The necessary experimental data are urgently required to perform a more in-depth analysis of the issue. Subsequently, future work should be on dedicating an axial stress–strain model for FRACC.

In terms of column’s axial capacity prediction, ML models such as the XGBoost performed more effectively as compared with conventional approaches. This is because of its superior capacity in estimating the axial compressive strength of FRP-strengthened RC columns. Furthermore, an ANN model was found efficient in predicting the axial capacity of FRP-reinforced concrete columns.

4.2.2. Compressive Strength of FRP-Confined Concrete

It is well known that axially loaded FRP-confined concrete exhibits considerable and better mechanical characteristics compared with unconfined concrete. With a view to predicting the ultimate state of FRP-confined concrete, Keshtegar et al. [61] developed a unique hybrid model that combines the response surface model (RSM) and SVR. Predictions generated by the suggested model have been compared with those obtained with six empirical models and two data-driven RSM and SVR models using a database of 780 circular column specimens. According to a statistical analysis, the suggested RSM-SVR model predicted the compressive strength and axial strain of the FRP-confined concrete more precisely than the current models.

Ilyas et al. [62] presented the use of a multi-expression programming (MEP) model to anticipate the compressive strength of CFRP-confined concrete. To examine the model’s performance, a thorough statistical analysis was performed. Furthermore, the findings and predictions of the provided model were validated by adding a parametric analysis, and the model’s dependability was compared with other experimental and theoretical models found in the literature. The suggested model was well trained to accurately anticipate the strength of structural elements strengthened with CFRP, as proven by parametric and statistical analyses.

Ilyas et al. [63] confirmed that GEP has the capacity to predict the compression strength of circular CFRP-confined concrete columns. Based on a huge and reliable database containing 828 data points, a new GEP model was created. Compared with other AI systems, such as ANN and the adaptive neuro-fuzzy interface system (ANFIS), only GEP has the ability and robustness to deliver output in the form of a simple mathematical relationship that is simple to utilize. It was revealed that the GEP models built for the planning and design framework of CFRP-confined concrete columns included the impacts of all relevant explanatory variables. The suggested model outperformed current models in terms of efficiency and accuracy.

Data-driven Bayesian probabilistic and efficient ML models including SVM, BPNN, and multi-gene genetic programming, were presented by Chen et al. [64]. First, a complete database encompassing 471 test results on the ultimate states of FRP-confined concrete cylinders was assembled from available literature, and the database’s quality was meticulously evaluated and examined. Then, using the Bayesian parameter estimation approach, an update mechanism was created to evaluate the key parameters in the current models and to modify the selected existing models accordingly. The accuracy of data-driven Bayesian probabilistic and ML prediction models might be readily improved by adding new test data to the existing database.

Moodi et al. [65] investigated the effectiveness of three different ML methods, including radial basis function neural networks (RBFNNs), multi-layer perceptron (MLP), and SVR, for predicting the ultimate strength of square and rectangular columns confined using different FRP sheets. The capability of ML to predict the compressive strength of concrete constrained by FRP was proved by comparing the ML-derived results with the experimental data, which were in very excellent agreement. MLP and RBFNN both produced accurate results, and they offered more accurate estimations for figuring out the compressive strength of FRP-confined concrete. Additionally, the outcomes demonstrated that the RBFNN approach performed poorly, with a wider gap in statistical indicators for training and testing specimens than the MLP method. Additionally, more thorough research is necessary to determine the compressive strength of the square and rectangular concrete FRP-confined columns. Within this scope, future research should be directed at the use of hybrid soft computational approaches (ANN methods with optimization algorithms) or novel techniques such as high-correlated variable creator machines, multiple Ln equation regression, and genetic programming.

Du et al. [66] investigated a technique for determining optimal parameter combinations for anticipating the confinement impact of FRP using hyperparameter optimization with Bayesian fine tuning. The proposed model’s Bayesian optimization (BO) and XGBoost regressor predictions for a database of 820 columns with a circular cross-section were contrasted with those of six empirical models and a non-optimized ML regressor of XGBoost. The new model (BO-XGB) more accurately predicted the compressive strength and axial strain of concrete confined by FRP when compared with the empirical models.

Berradia et al. [67] developed a new ANN model using the Group Method of Data Handling (GMDH) to predict the compressive strength of FRP-confined normal-strength concrete (NC) cylinders. To provide the maximum forecast accuracy, numerous factors were used. Through prior research investigations, a sizable experimental database of 313 FRP-confined NC cylinders was created. Using several statistical measures (root mean squared error, mean absolute error, and coefficient of determination R²) across the created database, an analysis of 33 alternative empirical strength models was conducted. The results of the study demonstrated that the proposed ANN model was capable of properly representing the compressive strength of FRP-confined NC cylinders, which could be utilized for future analysis and design of such members in the construction sector.

Cui et al. [68] used the normalized AlexNet-ELM and the sophisticated Red Fox optimization method to forecast the compressive strength of FRP-confined concrete in circular columns. The validity of the predictions was demonstrated through a parametric study, and the precision was assessed by an empirical versus theoretical comparison. A supplemental comparison was shown by taking into account the theoretical prediction obtained from the presented approach and the outcomes of the formulations employing important design codes. As a result, the approach used was customized to the FRP-confined concrete design and guaranteed a higher degree of precision in comparison with the available rivals.

The axial load-carrying capacity of elliptical-shaped sections of confined concrete reinforced with CFRP and steel tube was examined by Isleem et al. [69]. ANN modeling and finite element analysis were utilized. The findings demonstrated that the circular concrete section of the column, confined by steel tubes, could be used for a novel architectural type of construction. This study might be expanded to include eccentrical loading scenarios, different slenderness ratios, and other FRP types.

Sofos et al. [70] predicted the compressive strength of FRP-confined concrete specimens by applying ML tools based on experimental measurements. The material’s mechanical and physical properties had a relationship to the experimental measurements, which were delivered to a platform for machine learning. The experimental dataset was initially prepared using innovative data science approaches before moving on to the ML process. To forecast compressive strength, twelve ML techniques were used, with tree-based methods producing the greatest accurate results and obtaining coefficients of determination that were near unity. Ultimately, it was demonstrated that by carefully modifying experimental datasets and choosing the right algorithm, a quick and accurate computational platform was created. This platform was then generalized to avoid costly, time-consuming, and error-prone experiments and to provide a practical solution to issues in science and engineering.

The mechanical behavior of FRP-confined concrete was predicted by Tijani et al. [71] using the fundamentals of ultimate strength and strain. In this study, two ML approaches, ANN and GPR, were used to evaluate the observations from 627 datasets of FRP-confined concrete columns subjected to axial loading. The findings highlighted the value of applying AI approaches in structural engineering applications due to their amazing capacity to grasp multidimensional phenomena of FRP-confined concrete structures with ease, cost-effective computing, and superior performance over existing empirical models.

Existing confinement models, however, have poor predictability and cannot serve as a useful guide for real-world applications. A database with experimental data on 221 FRP-confined normal concrete cylinders was obtained from the available literature [72]. Then, a confinement model was created using GMDH approach. The prediction outcomes of these models were compared with nine other models and assessed with five all-inclusive metrics. The outcomes showed that the GMDH model can accurately estimate confined concrete compressive strength and ultimate strain.

The compressive strength of FRP-confined concrete was investigated and determined by Jamali et al. [73]. A large database of 1066 confined concrete cylinders with FRP sheets was collected. The calculation and assessment of the compressive strength of the aforementioned specimens using ML techniques were then described. The methods employed included the MLP artificial neural network, fuzzy neural inference system, particle swarm optimization (PSO), and kriging interpolation method. The findings demonstrated that, when compared with other models, the Kriging interpolation approach had the lowest error for evaluating compressive strength.

Sayed et al. [74] employed ML to calculate the axial compressive load of FRP-confined rectangular RC columns. The development of ML models, such as gradient boosting (GB) and RF, was accomplished by gathering datasets from previously published research. The produced models were then contrasted with pre-existing design-oriented models. The suggested ML models were found to be in strong agreement with the dataset test results. When comparing current design-oriented models to newly constructed ML models, the GB and RF regressors were closer in precision, and both techniques attained the lowest deviation values. The future task should be to develop the ML model using a larger database. Additionally, it is crucial to combine machine-learning-based models with other models (such as finite element models) to produce more reliable and precise models for engineering design.

A data-driven ML model was created by Li et al. [75] to simulate the compressive strength of glass fiber-reinforced polymer (GFRP)-confined RC columns and to analyze the significance and sensitivity of the parameters impacting the compressive strength. Data were collected from 114 sets of GFRP-confined RC columns from earlier studies. The findings indicated that the researchers’ model had a high coefficient of variation and a moderate estimate of compressive strength. When predicting the compressive strength of constrained columns, the BPNN had superior performance in terms of accuracy and resilience.

Analytical and ML models were developed to estimate the compressive strength (CS) of FRP-confined concrete cylinders by Kumar et al. [76]. The utilized ML models were SVM, optimized SVM, ANN, GPR, and optimized GPR. The improved GPR model has been discovered to be the best among all other models, according to the findings analyzed by the ML algorithms. Researchers, academics, and industry experts were able to forecast the CS of concrete cylinders enclosed in FRP according to the proposed work. Future research can be directed at the use of ML algorithms to estimate the CS and axial capacity of the corroded FRP-confined concrete cylinders. Furthermore, nature-inspired algorithms could be utilized to increase the accuracy and validity of anticipated models.

The use of ML algorithms in predicting the compressive strength of confined concrete improves the efficiency, accuracy, and resilience of these methods. Moreover, by generalizing these models, scientific and engineering problems may be solved in feasible, accurate, and practical approach. The generated ML algorithms used for predicting the capacity of FRP-strengthened RC columns are displayed in Figure 8.

4.2.3. Bridge Columns (Drift Ratio)

A mathematical expression utilizing ML-based symbolic regression was applied by Osman et al. [77] to evaluate the drift ratio limit states and corresponding strengths for circular RC bridge columns strengthened with hybrid reinforcements: external GFRP and internal steel layers. In concrete bridge columns, the interaction between the two materials increased corrosion resistance while preserving stiffness and elasticity. In particular, hybrid RC columns’ global and local responses under monotonic displacement-controlled loads were predicted using a validated fiber-based model. Consequently, the suggested expressions might be used for designing a hybrid bridge column.

4.2.4. Fire Exposure Influence in FRP Strengthened Members

The ANN was being creatively used to predict entire temperature profiles over time for fire-exposed circular reinforced concrete columns. This application was suitable for un-strengthened or strengthened columns with FRP and isolated with various types and thicknesses of insulating materials under variable fire exposure conditions. The maximum applied temperature was 800 °C. Furthermore, the temperature fluctuation with time for column constituent materials such as concrete, reinforcing steel, and FRP sheets was predicted. The input parameters used in data training are represented in Figure 9. The research findings demonstrated that the created ANN model had an overall accuracy of 85–90% in predicting the temperature of steel reinforcement, concrete, and FRP during fire exposure [78].

Naser et al. [79] proposed a methodology for creating, measuring, and testing various supervised ML algorithms against structural and fire engineering databases. Six algorithms were developed, including DT, Extreme Gradient Boosted Trees (ExGBTs), LGBT, RF, TensorFlow Deep Learning (TFDL), and Keras Deep Residual Neural Network (KDPNN). The presented investigation provided a foundation for a comprehensive framework that could be utilized for accelerating the adoption of ML in the structural and fire engineering fields. The study’s findings demonstrated that all chosen algorithms—with varied degrees of success—seem to accurately capture the structural and fire engineering phenomena under investigation. It indicated that structural and fire engineers may use raw algorithms rather than creating complicated ML models or going through rigorous programming exercises. The lack of ML workshops in structural and fire engineering courses further suggested that difficulties brought on by engineers’ historically limited understanding of ML coding could be readily resolved.

Bhatt and Sharma [80] designed a data-driven deep neural network (DNN) to evaluate the fire resistance time of reinforced concrete beams strengthened with FRP. Both a scaled and an unscaled dataset were used to train the model. For this, a large dataset of concrete beams that had been strengthened with FRP considering various geometry, insulation configurations, applied loads, and material properties was generated. After significant hyperparameter tuning and a ten-fold cross-validation technique, the DNN structure was chosen. The DNN model offers a relatively accurate assessment of the fire resistance of FRP-strengthened concrete beams. The analytical results showed that the thermal characteristics of insulation were crucial in influencing the fire resistance of FRP-strengthened concrete beams.

4.3. Reinforced Concrete Slab

Reinforced concrete flat slabs are considered one of the most popular floor systems that are widely used in buildings due to their essential advantages, such as ease and speed of construction. The use of these elements offers the flexibility to modify the functional needs of buildings in terms of increasing or changing the living spaces, as modern architectural designs may require. However, the vulnerability of flat slabs, when poorly designed, could lead to deadly catastrophes. One of the most dangerous failure modes in this type of slab is punching shear, which is a brittle two-way shear failure that initiates at the slab–column interface. Within this scope, various ML algorithms were developed and implemented to establish mathematical equations to predict the punching shear strength of RC slabs considering various parameters with relatively high accuracy [81,82,83,84,85,86]. This research revealed that 6.5% of the reviewed articles focused on the application of ML techniques for predicting punching shear strength.

Vu and Hoang [81] proposed a hybrid technique that combines least squares support vector machines (LS-SVMs) and the Firefly Algorithm (FA) to predict the punching shear strength of RC slabs. The input parameters considered were the slab effective depth, the shear span-to-effective-depth ratio, the aspect ratio of the column, and the concrete compressive strength. Then, the results were compared with those of ANN. It was found that the LS-SVM technique yielded highly accurate results, outperforming ANN with R² values of 0.9 and 0.97 for the training and testing processes, respectively. Recently, modified compression field theory (MCFT) was used to derive a hybrid ML model that involves XGBoost and SHAP methods to investigate further the punching shear capacity of FRP-RC slabs [86]. The resulting symbolic regression MCFT (SR-MCFT) model was superior in predicting the punching shear capacity of FRP-RC slabs as compared with existing empirical models.

Furthermore, three different ML algorithms, namely, ANN, RF, and SVR, were employed to predict the punching shear strength of RC slabs [82]. The influencing variables considered included the concrete compressive strength, the thickness and type of FRP, and the slab span-to-depth ratio. The results showed that all three algorithms were able to accurately calculate the punching shear strength of FRP-reinforced concrete slabs without the need for additional shear reinforcement. Instead of ANN, Truong et al. [85] conducted a similar study using the XGBoost technique to predict the punching shear strength of FRP RC-slabs, but with no shear reinforcement. The effect of slab depth, shear span-to-depth ratio, and concrete compressive strength was explored. The XGBoost model resulted in a lower discrepancy as compared with the SVR- and RF-based models. It was also shown that effective slab depth was the most influential parameter on prediction performance.

The FRP slab–column connection was comprehensively investigated using various ML algorithms, with an emphasis on punching shear failure [83]. Five different algorithms, namely, ET, GRP, LR, RDT, and SVM, were implemented to compare the accuracy of each model considering the effects of slab depth, column dimensions, flexural reinforcements, and concrete compressive strength. When compared with previously published datasets, the ET model had the highest accuracy in predicting the punching shear strength of FRP-RC slabs, with the slab’s effective depth being the most influential parameter. It was also shown that the developed models were superior to the existing design equations, indicating that the proposed models provide a reliable alternative to the existing mathematical solutions.

To demonstrate the accuracy of ML, another recent study on the punching shear strength of FRP RC-slabs was conducted using various ML algorithms, including ANN, SVM, DT, and AB [84]. The considered variables were the shape and size of the column’s cross-section, slab effective depth, and concrete compressive strength. The AB model was found to be the most accurate model among the adopted algorithms, with nearly zero deviation from the experimental results. However, all ML models showed enhanced performance when compared with the existing empirical equations.

In summary, implementing various ML algorithms showed different levels of prediction improvements. Among these algorithms, ensemble models such as XGBoost showed exceptional predictive accuracy, offering a promising alternative to existing empirical and analytical solutions.

4.4. Beam–Column Joints

The results showed that only 1% of the reviewed studies focused on the application of ML techniques to beam–column joints. Kisswani and Alubaid [87] developed an equation for evaluating the structural behavior of RC beam–column joints under cyclic loads using AI through deep learning to achieve the goal of minimizing errors. A regression model based on Python Anaconda, Jupyter Notebooks, and TensorFlow was utilized to apply deep learning. The study evaluated the structural behavior of RC joints strengthened with ferrocement and CFRP under cyclic loading, including stress, energy dissipation, stiffness degradation, ductility, displacements, tensile damage, compressive damage, plastic strain, and plastic dissipated energy density. Finally, a prediction equation for these aspects was developed, and the best reinforcement details with the minimum errors were suggested.

4.5. Bond Strength

Another critical problem with FRP-strengthening RC structures is the bond behavior at the FRP–concrete interface. Estimating the bond behavior and strength of such RC structural members is critical for the design and practical implementation of concrete structures. Over past decades, tremendous efforts have been directed toward developing analytical and empirical models that predict the bond strength between FRP and concrete. Recently, ML has emerged as a novel alternative technique for developing enhanced solutions that calculate the bond strength between FRP bars or sheets and concrete. The results showed that 22% of the reviewed studies discussed the application of different ML techniques for FRP–concrete bond strength. Within this scope, many ML models have been created to compare the feasibility of various algorithms [88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103].

With the available datasets from the literature, most of the aforementioned studies employed supervised learning techniques due to their capability of accurately modeling complex problems that involve a significant number of input parameters. Su et al. [90] developed three different ML models, namely, ANN, Multiple Linear Regression (MLR), and SVM, to predict the interfacial bond strength between concrete and FRP laminates. Two large single-lap shear datasets (122 and 136 test results) of concrete prisms strengthened externally with FRP sheets were adopted to train the developed models. With regard to forecasting precision, it was shown that SVM had the least discrepancy among the three models, resulting in more reliability as compared with MLR and SVM. A similar study involving ANN, MLR, support vector machine regression (SVMR), GPR, and regression tree (RT) was conducted to predict the bond strength between FRP bars and concrete [93]. When compared with the experimental results from the literature, the GPR model was found to yield the most accurate results, followed by ANN, MLR, and SVMR. Recently, Zhang et al. [103] formulated the bond capacity of externally bonded FRP sheets using five ML techniques, namely ANN, SVM, GBDT, DT, and XGboost. A large database of 1375 direct shear test results were used to confirm the models’ correctness. All ML models showed enhanced performance as compared with empirical solutions from the literature, with the XGboost algorithms having the best prediction outcomes. Another supervised ML method, namely GMDH, was implemented by Hamze-Ziabari and Yasavoli [88]. The model was developed with reference to 342 single pull-out tests’ results from the literature. It was concluded that GMDH outperformed existing empirical equations in terms of accuracy and safety. In addition, ANN was used along with MLR and ANFIS to determine the shear bond strength of FRP–-concrete connections under water immersion conditions [94]. The ANN approach was superior to the other models in computing the interfacial bond strength of FRP-strengthened joints under water immersion conditions.

In order to identify the cohesive zone parameters at the FRP–concrete interface of finite element models (FEM), an ANN-based algorithm was developed [91]. Many FEMs were generated to establish load versus displacement curves, which were used to train the ANN model. It was shown that the ANN-based model was able to accurately capture the FRP–concrete interface behavior with the ability to interpolate among the training datasets, indicating the feasibility of the model in such applications. Yuan et al. [98] proposed a new method to determine the cohesive parameters of the FRP–concrete interface and bond-slip behavior using ten regression models, including, but not limited to, the LR, SVR, DT, and Catboost regressors. With the use of FE modeling and Bayesian hyperparameter optimization, the Catboost regressor had the highest prediction accuracy.

Furthermore, many ensemble models were developed for the prediction of FRP-to-concrete bond strength. Chen et al. [92] applied an ensemble ML algorithm, namely gradient-boosted regression trees (GBRTs), to formulate a general solution for FRP–concrete bond strength. With the use of a comprehensive dataset including 520 test results, the developed solution had the highest accuracy as compared with existing empirical models. In a study conducted by [96], the CatBoost algorithm was utilized to generate a model capable of predicting the bond capacity at the FRP–concrete interface. The model was compared with other ensemble methods, including GBRT and RF. Based on a dataset of 855 single-lap shear test results, it was proven that the CatBoost algorithm was able to effectively predict the bond strength between FRP and concrete and was superior to the other ensemble models. The durability and failure modes of FRP and concrete connections under moisture conditions were measured using ANN, SVM, DT, and ensemble models [94]. The developed models were trained and tested using 429 durability test results collected from the literature. ANN-based models had the most accurate prediction of the bond strength. However, the ensemble model was superior to the other models in predicting the failure mode of FRP–concrete connections.

Amin et al. [100] investigated the bond strength of FRP sheets attached to concrete prisms through grooves. Three ensemble models were identified: RF, XGBoost, and LIGHT GBM. Using a dataset of 136 single-lap shear test results, LIGHT GMB yielded the most precise predictions, followed by XGBoost. Similarly, Alabdullh et al. [99] employed ANN, ELM, GMDH, MARS, LSSVM, and GPR to develop a new ensemble model called HENS for the prediction of the bond strength of FRP-strengthened concrete prisms with grooves. The ensemble model was superior to the other developed algorithms in terms of prediction accuracy.

On the other hand, the ELM was applied to study the bond behavior at the interface between concrete and FRP bars [101]. In this study, a new optimized model, namely a particle swarm optimization-based ELM model, was proposed. Based on the results of 222 specimens from the literature, the results revealed that the proposed model yielded highly accurate predictions, therefore improving the prediction performance of the bond strength between concrete and FRP bars as compared with the original ELM models.

In addition, a combination of ANN and Artificial Bee Colony (ABC) was proposed as a new ML method to estimate the interfacial bond strength between concrete and FRP [89,95,104]. In these studies, the results of large datasets of single and double-lap shear tests were collected from the available published articles to train the developed ABC-ANN models. These hybrid models were found to be more robust and yielded accurate results as compared with the ANN algorithm and existing design codes. However, when compared with ICA-ANN and GPR, the developed ABC-ANN models had a slightly higher discrepancy with respect to the test results collected from the literature. This could be attributed to the limitations of the input parameters used in the training process of the models [90]. Furthermore, ANN was fused with new population-based algorithms, namely bald eagle search (BES), dynamic fitness distance balance-manta ray foraging optimization (dFDB-MRFO), and RUN, to generate robust hybrid models [97]. Among the developed models, RUN-ANN was superior in predicting the bond strength at the FRP-concrete interface.

Based on ML techniques, Rui [105] developed some data-driven models to evaluate the intermediate flexural crack-induced interfacial debonding strain (IC debonding) of RC beams reinforced by FRP. The ML techniques are needed for the suggested model. It was discovered that the BP model does a good job at predicting IC debonding strain. The back-propagation (BP) data-driven model, on the other hand, is very difficult to converge and is prone to local minimums, which has an adverse impact on the model’s accuracy. To improve it, the sparrow search algorithm (SSA) was suggested. According to the findings, the neural network optimized by SSA with the lowest relative error is the most accurate in foretelling IC debonding strain.

Plate end (PE) debonding is a common issue with RC beams strengthened with FRPs. Hu and Li [106] established ML algorithms to predict the PE debonding of FRP-strengthened RC beams, considering the incredibly complex nonlinear relationship between the parameters and the PE debonding. These algorithms include LR, ridge regression, RF, decision tree, and NN improved by the sparrow search algorithm. The results indicate that the model developed in this paper has two issues: first, the model’s prediction accuracy can be improved due to the dataset’s uneven distribution of parameters, and second, the model is overly complicated due to the inclusion of too many parameters. In the future, additional data must be gathered, and the model’s parameters must be simplified.

Gu and Unjoh [107] present a method for utilizing infrared thermography to identify delamination in concrete structures jacketed with CFRP. The concrete specimens under consideration were examined in a variety of weather conditions, including winter, summer, sunny, and rainy. In this work, infrared thermography was used as a tool for image processing. The analyzed parameters included the specimen depth, size, surface cover mortar, and water content in the delamination void. The approach recognized delamination zones via boundary recognition based on changes in surface temperature fluctuations over time. The suggested approach was shown to be more effective and precise at detecting delamination than visual evaluations based on thermal images. Furthermore, following image processing using the suggested methods, a few delaminations that were invisible on thermal scans were identified. Additionally, the testing window and data collection intervals had a big impact on how accurate the results were. For future research, the authors recommended that a library of experimental data based on new specimens could be used to build a deep learning model.

As compared with traditional methods, it was evidenced that the use of ML showed significant improvement in the prediction accuracy of complex engineering problems like the FRP-concrete bond strength. In general, advanced techniques like ELM and ANN-ABC demonstrated remarkable and robust results, indicating the critical role of ML in the field of FRP strengthening. On the other hand, ML has also been successfully applied to specific challenges, like identifying delamination in FRP-strengthened concrete specimens using infrared thermography, indicating the potential use of ML to target more specific engineering problems. The ML algorithms utilized in predicting the bond strength at the FRP–concrete interface are shown in Figure 10.

4.6. Materials

The results revealed that 6.5% of the reviewed studies focused on the application of ML techniques to FRP bond strength. The strength prediction of fiber-reinforced concrete, the post-fire behavior of construction materials, measuring strain in pre-stressed FRP, and FRP defect detection in composite systems are the main sub-themes.

4.6.1. Strength Prediction of Fiber-Reinforced Concrete

Basalt FRP geopolymer composites are seen as a long-term solution to the structural and environmental difficulties associated with traditional RC constructions, which include increasing carbon dioxide emissions and structural degradation caused by corrosion. The limited application of FRP-based geopolymer composites is due to the inaccuracy of the current design standards in predicting flexural capacity and strength properties. Rahman and Al-Ameri [108] investigated the flexural strength of basalt FRP-reinforced self-compacting geopolymer concrete beams under varied exposure situations using ANN-based models. Three different algorithms namely, Levenberg–Marquardt (LM), Bayesian regularization (BR), and Scaled Conjugate Gradient (SC) were used to create three distinct ANN models. Experimental data from a detailed investigation of beam specimens exposed to one-year ambient and marine environments were employed to train and evaluate the developed models. The ANN prediction models showed strong agreement with the experimental results and higher prediction accuracy than the previous empirical and numerical models. For the flexural design of FRP-reinforced structures, current design predictions based on ACI formulations must take degradation and environmental variables into account for long-term performance estimations. Future research that generates hybrid models by combining the current design standards with ANN tools should promote a wider use of innovative reinforced geopolymer concrete composites. According to Abdellatif and Raza [109], there have been few investigations on the mathematical estimations of the compressive strength (CS) of embedded glass fiber-reinforced polymer (glass-FRP) members. The authors applied ANNs and mathematical modeling to predict the CS of members made of glass-FRP and normal-strength concrete (glass-FRP-NSTC). The authors used MATLAB’s curve-fitting and general regression techniques to construct a novel mathematical equation to estimate the CS of members made of glass fiber and NSTC. To ensure optimal estimations, the recently proposed ANN equation was calibrated for various hidden layers and neurons. Comparing the recommended equations’ estimations to those of the equations that are available in the literature, the suggested equations showed a strong correlation among themselves and provided correct estimates.

4.6.2. Post-Fire Behavior of Construction Materials

Knowledge of post-fire material characteristics (e.g., compressive strength, tensile strength, yield strength, and Young’s modulus) is essential for assessing the residual capacity of buildings after a fire. Naser and Uppala [110] initiated a study that illustrates a method to derive residual material models for a range of construction materials, including normal strength concrete (NSC), high strength concrete (HSC), ultra-high-performance concrete (UHPC), mild steel (MS), stainless steel (SS), cold formed steel (CFS), high strength steel (HSS), and glass fiber-reinforced polymer (GFRP). In this method, two ML approaches, ANN and GA, were combined in a hybrid fashion to create material models that were especially tuned to track the behavior of building materials after a fire. In summary, ML had the potential to be a key tool for realizing generic material models. In order to improve post-fire investigations, recent research called for updating fire assessment techniques as well as generating innovative representations of the residual characteristics.

4.6.3. Measuring Strain in Pre-Stressed FRP

The measurement of strain levels during pre-stress applications is often difficult and time-consuming. A significant advancement in the subject is the creation of rapid approaches for measuring the pre-stressed application of laminates. Recent research suggested employing ML and deep learning applications, such as benchmark algorithms, to measure the CFRP strain level. Valença et al. [111] presented benchmarks of contact-free design for monitoring the strain level of pre-stressed CFRP laminate based on computer vision to develop a system that might be economically practical, easy-to-use, automated, and accurate. The architecture was energized by digitally deformed synthetic images created from a low-resolution camera. It was crucial to note that the described architecture was automated, measures directly on the laminate surfaces, and was cost-effective, which allowed it to be extensively employed in the application of pre-stressed FRP laminates. A theoretical model was created by Yan et al. [112] to examine the long-term prestress loss of reinforced elements along with the adhesion behavior of the CFRP–concrete interface under natural exposure scenarios. The goal of this article was to further the development of CFRP reinforcement technology by investigating the features of engineering geomechanics based on ML techniques. The findings indicated that theoretical research has to be strengthened and successfully applied to practical problems. Informed by the reliability hypothesis, workable rules and standards are created for direct practice.

4.6.4. FRP-Defect Detection in Composite Systems

The acoustic-laser approach has proven to be an excellent non-destructive testing (NDT) methodology for finding defects in composite structures. The approach is especially beneficial for detecting near-surface faults in fiber-reinforced polymer-bonded concrete by shaking the material with acoustic stimulation and monitoring the vibration signals with a laser beam. However, just like other vibration-based measuring techniques, the accuracy of the acoustic-laser approach depends on the sample rate used during the measurements. So, the measurement results with continuous or random missing data were recovered using the K-singular value decomposition (K-SVD) dictionary learning technique [113]. The recovered findings included increased measurement accuracy, and the rebuilt signals demonstrated good agreement with the original measurement results without missing data. The usefulness of the K-SVD dictionary learning approach for up sampling and missing data recovery in acoustic laser technology has been confirmed.

5. Parameters

Based on the comprehensive review provided above, most of the datasets used for the development of the ML models involve a large number of parameters. These parameters have a significant influence on the models’ outcomes. Some of these crucial parameters are as follows: (1) the type of FRP material being utilized—such as carbon, glass, or aramid fibers—as well as its characteristics, such as stiffness, durability, and tensile strength; (2) the geometrical-related parameters, such as the slenderness ratio of the columns, size of members, and bond and embedment lengths at the FRP–concrete interface; (3) the mechanical parameters, which include material strength and elastic modulus; and (4) the physical parameters, such as reinforcement ratio (steel and FRP) and FRP dimension. The parameters investigated for each particular application are presented in

Table 2. Parameters investigated and used ML models.

Applications	Studied Parameters	Most Used ML Models
Beams	Web width, effective depth of the section, span-to-effective depth ratio, the compression depth, maximum aggregate size, compressive strength of concrete, elastic modulus of concrete, elastic modulus of steel rebars, tensile strength of rebars, steel reinforcement ratio in flexure and shear, modular ratio, type of FRP bars used as longitudinal reinforcement (i.e., CFRP, AFRP, GFRP, BFRP), type of FRP bars used as shear reinforcement, FRP thickness, FRP longitudinal reinforcement ratio, FRP shear reinforcement ratio, elastic modulus of FRP longitudinal reinforcement bars and FRP stirrups, ultimate tensile strength of FRP longitudinal reinforcement bars, FRP shear reinforcement bars strength, applied loading conditions.	ANN, DNN, XGBoost, RF, ELM, GEP, EL, GBDT, and SVM.
Columns	Slenderness ratio, gross cross-sectional area, type of cross-section (circular or rectangular), type of concrete (light-weight or normal-weight concrete), type of aggregate, compressive strength of concrete, type of composite material used in the longitudinal and/or transverse reinforcements (i.e., GFRP, BFRP, and CFRP), longitudinal reinforcement ratio, elasticity modulus of steel, ultimate strength of steel, configuration of transverse reinforcement (i.e., spirals or ties), spacing of the transverse reinforcement.	ANN, LR, RF, XGBoost, SVM, RSM-SVR, GPR, and Bayesian optimization.
Slabs	Slab length, slab width, slab depth, concrete compressive strength, steel yield strength, steel reinforcement ratio, fiber cross-section, fiber tensile strength, bond strength, reflected impulse, reflected pressure, punching shear strength, slab type: one-way/two-way, FRP configuration.	ANN, SVR, RF, GPR, SVM, XGBoost, and DT.
Bond Strength	FRP elastic modulus, FRP tensile strength, FRP type, FRP thickness, FRP width, stiffness of FRP, bond length, concrete compressive strength, cross-section width, surface texture, groove width, groove depth, reinforcement diameter, reinforcement position, concrete cover, embedment length, the presence of transverse reinforcement.	ANN, GPR, SVMR, RT, MLR, ELM, SVM, GPR, and GBRT.
Materials	Different concrete materials: NSC, HSC, UHPC, different steel materials: MS, HSS, CFS, SS, different FRP materials: GFRP, CFRP, and temperature.	ANN and CAFM.

. Figure 11 shows the analysis and classification of the most frequently used parameters.

6. Discussion

The significance of using various ML techniques to predict the failure level of various RC members strengthened with FRP is discussed in this section. With the intention of providing the current state of research and based on the comprehensive review presented above, great potential is demonstrated for the use of ML to achieve accurate and robust solutions in the field of FRP-strengthening RC structures. When compared with existing empirical solutions and design codes, the applications of ML have significant improvements in the resistance prediction accuracy of all structural members strengthened with FRP, including the flexural and shear strengths of beams, axial capacity of columns, punching shear in slabs, and bond strength at the FRP–concrete interface. The consistent and significant improvement in prediction accuracy illustrated in this review addresses several limitations present in the current design codes.

Furthermore, the supervised techniques have garnered much attention in the field of FRP application to concrete members. In particular, the majority of the aforementioned literature has utilized regression algorithms including LR, NN, RF, DT, KNN, and BA to develop a mathematical model to estimate the resistance of a structural member. Generally, the boosting algorithms (BAs) are superior in terms of prediction accuracy and reasonable development costs. Other ML algorithms, however, have also shown great performance in predicting the behavior of FRP-RC members, especially when a wide range of parameters is considered. For example, ANN models have been frequently employed to predict the strength of FRP-RC members using extensive datasets that cover a large scale of input variables. This is due to the nature of the ANN models’ flexibility, non-linearity, scalability, and ability to learn from extensive datasets. In addition, the use of hybrid ML models also statistically enhanced the precision of predicting the members’ strengths.

On the other hand, the use of ML in the field of FRP strengthening still has some drawbacks that are limiting its application among the research community. These drawbacks are mainly related to the algorithm selected and the dataset used in the development process. First, the selection of the most reliable and cost-effective ML algorithm remains a challenge to this date. Based on the research provided above, many ML algorithms have resulted in highly accurate predictions. However, some of the reported outcomes are contradictory. For example, some studies have shown that ensemble models are superior to individual models for a particular application, while others have reported that both techniques have similar prediction performance. Considering every algorithm is developed differently, detailed and comprehensive studies need to be conducted to compare all algorithms for a particular application in the field of FRP strengthening. It should be noted that the selection of the most accurate and cost-effective model for each application is essential to consider how a specific issue is solved and to implement such solutions in the design codes for practical applications.

The second major challenge is related to the data used to develop these ML models. Since the topic of FRP strengthening RC members is relatively new in the field of structural engineering, the available experimental data in the literature can be considered somehow inadequate. This can be clearly seen from the reported results of the research above, where the change in the dataset size can lead to inconsistent outcomes. In addition, some of the important parameters, which significantly influence the strength of a specific member, were only investigated once or twice in the literature. Although the effect of such parameters such as the size of the member is not necessarily considered in the current design codes, it could affect the performance of the model drastically. With this in mind, statistical or finite element analysis studies should be performed to identify the most influential variables of a particular application under various conditions, which could be used in the development of a comprehensive ML model.

It should also be noted that the vulnerability of deteriorated buildings could result in a significant number of casualties, as observed in the aftermath of severe earthquakes lately. Therefore, designers and practitioners who are knowledgeable about the use of ML should approach the development of such models with caution.

7. Conclusions

Sustainable solutions in the building construction industry, like the use of FRP, have emerged as a new method in retrofitting applications in the last two decades. The existing design guidelines for FRP-strengthening RC structures are developed using empirical solutions, which are mainly based on an inadequate number of studies, limiting the accuracy of the predicted results. Therefore, the use of innovative and efficient prediction tools such as ML, which could be developed using comprehensive datasets covering a wide range of parameters, has become essential. Thus, this paper presents a state-of-the-art review of the advances in implementing ML in FRP strengthening applications. The current study covers a broad range of ML applications in the field of FRP strengthening, including the flexural and shear strengths of RC beams, the confinement and compressive strength of columns, the flexural capacity of slabs, and FRP–concrete interfacial behavior. With this intention, the outcomes of previous research, the significance of the implemented techniques, the challenges, and the future directions are discussed in this review.

In a nutshell, the use of ML has shown great potential for achieving accurate solutions in FRP applications and strengthening RC structures. With respect to existing empirical solutions and design codes, the application of ML has a significant improvement in resistance prediction accuracy. Based on the research presented above, it is observed that supervised learning techniques, where input and output data are used in the development process, are the most favorable learning method due to their good generalization, interpretability, adaptability, and predictive efficiency. Furthermore, ensemble models developed by combining two or more individual algorithms have shown higher predictive capabilities when compared with single algorithms. However, some studies stated that both techniques have comparable prediction performance. Therefore, each model might be more appropriate for a specific application, considering that ML algorithms are built differently. In addition, optimization techniques such as Bayesian optimization and particle swarm optimization have shown a significant impact on the performance of the ML algorithm adopted. The selection of appropriate ML algorithms and optimization techniques is found to be more influenced by many factors, including the nature of the problem and the characteristics of the dataset.

On the other hand, the use of ML in the field of FRP strengthening has limitations that need to be addressed before the implementation of the design codes. The selection of an appropriate ML algorithm by means of efficiency and cost remains a major challenge in this field, as evidenced by the inconsistent outcomes reported above. Moreover, the literature still lacks the use of appropriate optimization techniques due to the relatively limited application of ML techniques. The second major challenge is the limitation of the available experimental data in the literature since the topic of FRP strengthening RC members is relatively new in the field of structural engineering. This can be clearly seen from the reported results of the research above, where the change in the dataset size resulted in different outcomes. In addition, more than 60% of the reviewed articles focus only on the application of ML in beams and columns, whereas the studies on slabs and beam–column joints are less than 10%. Finally, some of the crucial parameters that impact the strength of a structural member are only investigated once or twice in the literature. Such parameters are believed to influence the performance of the model considerably, such as the size of the member.

Based on the conclusions and challenges discussed in this study, the following suggestions are provided:

• Since every ML model has its own unique characteristics, detailed and comprehensive studies need to be conducted to compare all algorithms for a particular application and to select the most appropriate ML technique. Similarly, further research is necessary to determine the most suitable optimization algorithm for each application.
• Due to the complex behavior of FRP-strengthened RC slabs and beam–column joints, comprehensive investigations involving ML application in predicting the behavior of such members are needed to account for the key parameters affecting their ultimate performance.
• Statistical and finite element analysis studies that involve a wide range of parameters should be performed to identify the most influential variables of a particular application under various conditions, which could be used in the development of a comprehensive ML model.
• More studies could be conducted to estimate the FRP contribution for each application rather than the strength of the retrofitted RC members while considering a wide range of parameters.
• Since the use of ML showed great potential in improving the current design guidelines, it could be expanded to cover newer corresponding strengthening techniques like the use of textile-reinforced mortar (TRM) and shape memory alloys (SPAs) in strengthening RC structures.