Systematic literature review on the application of machine learning for the prediction of properties of different types of concrete

Rationale

The rationale for conducting this systematic review on machine learning applications in concrete is driven by the need to address the challenges and limitations of traditional structural analysis and design approaches. The construction industry heavily relies on concrete, which consumes significant amounts of virgin resources and plays a crucial role in building infrastructure. However, the conventional methods used for structural analysis often require time-consuming calibration procedures and struggle to handle severe actions with highly nonlinear behavior. Therefore, there is a need to explore alternative approaches that can reduce time and effort while improving accuracy and efficiency. Machine learning has shown promise in various fields, and its potential application in concrete structural engineering warrants investigation to identify its benefits and limitations.

The intended audience for this systematic review includes both structural engineering practitioners and researchers in the field of concrete construction. Structural engineers who are interested in exploring new approaches for structural analysis and design will find value in the overview of machine learning methods, principles, and available resources provided in this article. Researchers in the field of concrete and machine learning will benefit from the summary of existing studies, knowledge mapping, and identification of influential authors and nations. Additionally, professionals in the construction industry, including contractors, developers, and project managers, can gain insights into the potential benefits of machine learning in terms of cost savings, time efficiency, and labor intensity. Overall, this review aims to bridge the gap between traditional structural engineering practices and the emerging field of machine learning, providing a valuable resource for those seeking to incorporate ML methods into concrete applications.

Problem statement and research question

This is the most recent and state-of-the-art review on the application of machine learning techniques to predict the properties of different types of concrete. The goal is to conduct a literature review to summarize all the work done on the prediction of all the mechanical properties of concrete. This literature review will help future researchers to opt for the best algorithm for their concrete and later compare them with the work already done in this area.

Search engine and keywords

First, a set of keywords has been formulated which is given below to search the PubMed database for the relevant studies, then after removing duplicates and the inclusion and exclusion criteria discussed in Table 1 were applied to the rest of the studies which then resulted in narrowing the studies from 116 to 42 (Fig. 1). Then for the deeper search and in order to get the most possible and accurate results, the following keywords were also divided into different sets.

• (concrete technology) AND (mechanical OR durability OR compressive strength OR flexural strength OR modulus of elasticity OR tensile strength) AND (“computer vision” OR “neural network” OR “artificial intelligence” OR “pattern recognition” OR “machine learning”).

Eligibility criteria

Table 1 outlines inclusion and exclusion criteria for a study, likely related to the prediction of concrete properties using machine learning algorithms. These criteria are used to define the scope of the study and to determine which studies should be included in the analysis and which should be excluded. The inclusive criteria define the characteristics that studies must have to be considered for analysis (focus on concrete, use of machine learning, and publication in conferences or journals). The exclusive criteria define the characteristics that would lead to the exclusion of studies from the analysis (focus on non-concrete materials, use of methods other than machine learning, and lack of appropriate publication types). These criteria help ensure that the study’s scope remains relevant and focused on the specific research objectives.

Flowchart

Structural health monitoring

Civil constructions are subject to structural degradation as a result of their usage and environment. For the assurance of assure public safety and the in-service construction dependability, the Structural Health Monitoring (SHM) system is essential for early detection of structural problems. Dynamic response assessments separated at periodic intervals are used to monitor a component over time, damage-sensitive characteristics are recovered, and then the derived features are statistically examined to determine the present health condition of the system (Khoa et al., 2014). Long-span bridges, massive dams, and towering buildings are among the structures where the SHM system has been widely deployed, allowing for a seamless transition from time-based to situation management. Model-driven or data-driven techniques have both been used in recent studies in this area of interest. As a result of this method, it is possible to detect structural deterioration by comparing measured data to data generated by a computer model of the structure (typically based on finite element analysis (FEA)). Due to the repetitive examination of a simulation software model, this technique is computationally intensive (Jin, Cho & Jung, 2015; Salazar et al., 2015; Cha & Buyukozturk, 2014; Ranković et al., 2012). It is also possible that in actuality, a measurement simulation may not be available at all times or accurately represent the real structure’s performances in every case. Because of this, FEA findings are typically insufficient to accurately measure structural health. A strategy based on data rather than models generates a model via the use of observed data and then compares the model’s responses to those measured in order to discover damage. This method employs machine learning techniques, such as pattern recognition. It is becoming more possible to install large and dense sensor networks for SHM because to recent advancements in sensing methods, and wireless communication. As a result, continuous and real-time damage identification is made much easier with the data-driven method (Aldahiri, Alrashed & Hussain, 2021). To identify structural damage, machine learning algorithms are often used in conjunction with supervised learning, which relies on examples of both healthy and damaged data. Structural damage detection may benefit from the resilience and efficiency of single machine learning method such as support vector machine, neural networks, and support vector regressions, as well as the genetic algorithms (GA). For various challenges in the SHM sector, hybrid approaches such as the multi-objective genetic algorithm (MOGA), neuro-fuzzy (NF), and wavelet neural network (WNN) have also been presented. All investigations proved the accuracy of machine learning-based models and their better performance over model-driven methods (Ranković et al., 2012; Taffese & Sistonen, 2017).

Properties of mix design concrete

It can be seen in Table 2 that so many researchers contributed to predict the mechanical properties of the concrete mix with different substances like fly ash, foundry sand, or rubber waste using ML algorithms. Concrete buildings are designed with mechanical qualities including compressive strength, elastic modulus, splitting tensile strength, and shear strength in mind. Predicting the compressive strength of concrete by linear or non-linear regression equations saves both time and money (Chou et al., 2014). Elastic modulus measurement is difficult and time-consuming. Stress–strain relations of cementitious materials under compression are often used to get this information (Atici, 2011; Behnood, Verian & Modiri Gharehveran, 2015). The compressive strength of concrete is typically used to estimate the splitting tensile strength of concrete because of its complexity, expense, and time-consuming nature. Based on experimental data, regression models for shear strength of RC components are also applied. In the past, the mechanical characteristics of concrete were evaluated using a set equation that was based on a small amount of experimental data and variables. They are only useful for describing the results of their own experiments used to calibrate them. The model coefficients and the equation’s form must be updated if the original data is changed. To determine fresh concrete’s mechanical qualities, standard models may not be appropriate since the link between components and concrete characteristics is particularly nonlinear for certain concrete kinds. A widely agreed-upon mathematical model is also difficult to come by. A concrete structure’s long-term performance may be evaluated by looking at its dry shrinkage, another important feature of concrete. Several empirical equations for shrinkage estimation have been developed in various codes such as ACI and CEB throughout the last five decades. Dry shrinkage in concrete is affected by a variety of parameters, including its composition, the size of the specimen, and the quality of its ingredients. Using these calculations may be problematic in certain situations. Components and their relative proportions are determined in order to manufacture concrete that fulfills required strength, workability and durability at a low cost while yet delivering a high quality product. As an extension of previous practice, concrete mix percentage algorithms are typically available in the form of empirical formulae or tables. As a consequence of this uncertainty, typical methods for determining concrete mix proportions are a trial-and-error exercise, which results in higher expenses as well as more time (Nazari & Sanjayan, 2015). Modeling concrete characteristics and mix design accurately and reliably may save time and money by providing engineers with the information they need. To circumvent the limitations of standard empirical regression models, machine learning methods have been used to represent these features. Construction of accurate and effective models for predicting the characteristics and mix design of several kinds of concrete, including fiber-reinforced polymer (FRP) concrete have been done by using machine learning techniques. Many machine learning methods are used in these investigations, including neural networks, genetic programming, fuzzy logic, support vector machines, and fuzzy inference systems (FIS). Machine learning approaches have been shown to be a strong tool for evaluating tangible qualities, regardless of the complexity, incoherence, or incompleteness of the data used. They are also a superior alternative for deciding on the right quantities of materials in concrete mixtures to achieve the appropriate strength and rheology (Chou et al., 2014; Atici, 2011; Behnood, Verian & Modiri Gharehveran, 2015; Dantas, Batista Leite & de Jesus Nagahama, 2013; Nazari & Sanjayan, 2015; Mermerdaş & Arbili, 2015; Tang et al., 2015). Reducing trial mixes results in an ecological and cost-effective mix design method.

Artificial neural network

Parallel processing occurs in the brain’s neural network, which is a web of linked neurons that sends signals back and forth to process information. ANNs are a cutting-edge analytical technique that mimics the way the human brain thinks. Similar to other DoE approaches that take in numerous factors to forecast the response variable, ANNs may be employed mathematically to analyze multiple inputs and generate an output (Sharma, Rai & Dev, 2012). The input, hidden layer, and output layer are all parts of the ANN’s mechanism. It is here where data is entered. The output layer processes the data and provides the result via a system of connection weights. The inputs are fed into the process, and the process concludes with the output. A technique known as backward propagation is used to reduce the overall weight of the network’s connections. The discrepancy between the anticipated value and the actual value is believed to alter and change the mechanism of the hidden layer. It is important to understand the benefits and downsides of ANNs (Behnood & Golafshani, 2018). Due to its processing, errors may be tolerated, and complicated non-linear relationships between variables can be solved with ease using data analysis. ANNs have a distinct edge over pre-programmed computational models since they are able to learn from their own mistakes. It is also possible to overfit the data supplied by ANNs because of the intricacy of their solution (Sharma, Rai & Dev, 2012; Mohammed, Khed & Nuruddin, 2018).

Concrete compressive strength may be predicted using ANNs, which have a greater number of variables than previous DoE approaches. Analyzing many concrete experiments that all employ the same looking to upgrade is a unique use of ANNs thanks to their enhanced processing capability. Gupta (2013) who collected 32 data points from ten different publications on nano-silica-containing concrete, came up with an exact model for 28-day concrete compressive strength without having to do any experiments. Additionally, Asteris & Mokos (2020) utilized non-destructive test results from a thesis to train ANNs on 209 data sets to estimate concrete strength. Noorzaei, Hakim & Jaafar (2007) and Santosa & Purbo Santosa (2017) did a similar study utilizing the elements of concrete as variables and reached the same outcome. In terms of precision, regression analysis, particularly multiple non-linear regression, falls short in comparison to ANNs, as shown by the R2 value. When it comes to modeling self-compacting concrete, research found that the results of MLR outperformed those generated by ANNs. ANNs function best when given more data, and the low quantity of data in the study (i.e., 15) may account for this. The R2 score alone should not be utilized to choose the optimal model. The root mean squared error (RMSE) of the ANNs model was much lower than the other models in another experiment on recycled aggregate concrete.

Comparison and motivation of literature review

The PRISMA based methodology adaption for this systematic literature review has been taken from Zahid et al. (2022). This literature review is unique because it systematically summarizes the current state of research on the application of machine learning in the concrete industry, with a focus on structural analysis and design approaches. The review provides a comprehensive overview of the potential of ML to replace empirical models and reduce the time and effort required in the industry. It also provides an overview of ML methods, principles, access codes, libraries, and datasets that can be used by practitioners and researchers to develop their own ML models. Additionally, this review identifies the most active locations and influential authors in researching ML applications for concrete, which could facilitate future collaborations and sharing of novel ideas and approaches among academics. The statistical and graphical representation of contributing authors and nations can be useful for researchers and practitioners in identifying potential collaborators and networking opportunities. Overall, this review provides a valuable resource for researchers and practitioners in the concrete industry who are interested in exploring the potential of ML to improve their work. The systematic approach used in this review ensures that the information presented is comprehensive and unbiased, making it a valuable resource for anyone looking to learn more about the application of ML in the concrete industry.

Future trend

The great degree of accuracy in actual and predicted outcomes demonstrates the significance of these techniques in civil engineering. It is becoming increasingly common to use supervised ML techniques since they provide accurate outputs and reduce the amount of physical labor and overall project expense. In addition, it is vital to conduct laboratory experiments to compare the results of machine learning algorithms. In order to compare the results of different machine learning algorithms, it is also possible to alter or add input factors, such as the number of data points and the kind of material used, size of specimens, ambient conditions, curing settings, and data loading rate. For the sake of comparison, a variety of machine learning approaches may be used, including ANNs, SVMs, and boosting (Shang et al., 2022). Databases were used to calculate the compressive and split tensile strengths. As an alternative, additional input parameters and increasing the database may produce the required results. Silica Fume Concrete (SFC) is compressive and split tensile strength models have been created in this work. According to statistical characteristics, these models were able to accurately and reliably estimate SFC intensities. However, by using the same modeling parameters, MLPNN, ANFIS, and GEP models may be used to forecast concrete qualities including numerous different concrete ingredients. Based on input parameters, these models will be changed and the outcomes anticipated are largely dependent on the database used. The whale optimization algorithm, ant colony optimization, and particle swarm optimization are just a few examples of heuristic techniques that may be utilized in combination with machine learning to get optimum results. They may then be compared to this study’s methods. The upgraded and improved version of GEP is known as multi-expression programming (MEP). GEP’s limitations may be overcome via MEP analysis. To put it simply, MEP is given more attention when the complexity of the target expression is uncertain. There are exceptions, erroneous expressions, and even division by zero that can be handled by MEP. There are no infertile learners in the next generation since the gene is responsible for causing exceptions and then changing to an arbitrary terminal symbol. While MLPNN and ANFIS were used for the prediction of results, single learners were utilized in this study to anticipate results. Many different sub-models are built, and statistical parameters are used to pick the best one. This is known as an ensemble ML approach (Imran et al., 2022).