Data-driven model for ternary-blend concrete compressive strength prediction using machine learning approach

Highlights

• LSSVM using CSA for optimization was proposed in estimating the concrete compressive strength.

• The inclusion of all input variables provided a better representation of the developed model.

• LSSVM using CSA to optimize the hyperparameters improved the performance of the model.

• Sensitivity analysis revealed a substantial correlation between the input and target variables.

• The LSSVM-CSA model showed superior robustness and performance when compared with other studies.

Abstract

Ternary-blend concrete is a complex composite material, and the nonlinearity in its compressive strength behavior is unquestionable. Entirely many models have been developed to accurately predict the ternary-blend concrete compressive strength, such as ANN, SVM, random forest, decision tree, to mention but a few. This study underscores the better predictive performance and successful application of the least square support vector machine (LSSVM), a machine learning model for predicting the compressive strength of ternary-blend concrete. Coupled simulated annealing (CSA) was applied to the LSSVM model as an optimization algorithm. In addition, the genetic programming (GP) model was used as a benchmark model to compare the performance of the LSSVM-CSA model. The predictive performance of the LSSVM-CSA was compared with that of some of the proposed models in well-known studies where the same datasets were used. The model proposed in this study outperformed other studies, yielding an R² value of 0.954.

Introduction

Similar to every brittle material, the primary factor governing the compressive strength of concrete is the porosity. While there are other important characteristics of concrete, such as durability, impermeability, and volume stability, compressive strength has become universally accepted as the most important indication of concrete quality. Concrete compressive strength is affected by many factors, such as the quality of the raw materials (cement, supplementary cementitious materials, aggregates, water, etc.)., water-cement (w/c) ratio, coarse–fine aggregate ratio, age of concrete, proper vibration and compaction of concrete, relative humidity, and curing of concrete. The higher the w/c ratio, the greater the intergranular spacing between the cement particles and the greater the volume of voids left unfilled by hydration products after cement hydration. The coarse–fine aggregate (CA/FA) ratio is another influencing factor associated with the w/c ratio. The proportion of CA to FA influences the overall aggregate surface area, which affects the water demand of the concrete. With higher FA content relative to CA content, the water demand (w/c ratio) increases owing to the increased total aggregate surface area and vice versa if the FA content is lower. A higher or lower total aggregate surface area leads to a higher or lower w/c ratio, which in turn leads to a decrease or increase in the compressive strength, respectively. Since water is available for cement hydration, the compressive strength of concrete will continue, albeit at a continuously decreasing rate as it ages. Therefore, the age of concrete is synonymous with the degree of cement hydration provided that the concrete has not dried out or the surrounding temperature is not too low. In practical terms, a large portion of the concrete compressive strength would have developed after 28 d. To achieve a desirable compressive strength, curing is imperative for maximizing the concrete's potential for attainable compressive strength. Curing ensures the continuation of the hydration process in concrete, leading to continued strength gain. The hydration can continue due to the good preservation of moisture and temperature within the concrete for an adequate period of time.

Owing to the importance of the mechanical properties of concrete in reinforced concrete design codes and specifications, accurate prediction of these properties has become a concern. A significant concrete property is the compressive strength, which is usually measured after 28 days of standard curing [1]. Until recently, when computational intelligence, such as artificial neural networks (ANNs) [2], [3], [4], support vector regression (SVR) [5], [6], [7], [8], random forest (RF) [9], [10], [11], [12], etc., have been successfully applied in concrete research, estimating the mechanical properties of concrete has always been through linear and nonlinear regression methods [1], [13]. Other researchers have conducted studies to estimate the compressive strength using empirical methods where experimental data are calibrated, input ranges are limited, and constants required to describe the relationship between the inputs and the compressive strength are difficult to obtain [14], [15]. While some empirical models' effectiveness has been proven in some studies to save time and resources for future applications, multiple drawbacks remain in their development [16]. Several machine learning algorithms have been proposed as innovative predictive tools for predicting the compressive strength of concrete to overcome these weaknesses.

In concrete material studies, machine learning and data science have been used broadly in the optimization of concrete mixtures [17], [18], [19], fresh properties [20], [21], [22], [23], hardened properties [2], [5], [8], [10], [11], [12], [14], [24], [25], and durability properties [9], [26], [27], [28], [29] estimation of different types of concrete.

Using random forest, Silva et al. [30] presented a model to predict the compressive strength of concrete from 1030 experimental data with eight variables retrieved from Yeh's study [31]. The developed model outputs were compared with those from the support vector machine (SVM) and ANN. Comparing ANN and instance-based learning (IBL) or k-nearest neighbors (kNN), Dutta and Barai [32] explored the prediction efficiencies of these models in predicting the compressive strength of concrete using the mixture ingredients as features. Dao et al. [33] studied the sensitivity and robustness of Gaussian process regression (GPR) with five kernel functions and an ANN using a 500-run Monte Carlo simulation to estimate the compressive strength of high-performance concrete (HPC). The results revealed that the GPR with the Matern32 kernel function performed better than the ANN in predicting HPC compressive strength. Two models to forecast the compressive strength of HPC based on ANN ensembles (bagging and gradient boosting) with prediction accuracy enhancement through the coupling of discrete wavelength transform (DWT) with the ANN ensembles was proposed by Erdal et al. [34]. The DWT was concluded to be a useful tool for increasing the accuracy of the developed ANN ensembles.

Amid numerous other ML techniques, the least square support vector machine (LSSVM) has recently gained broader acceptance in concrete material science. [35], [36], [37], [38], [39]. With a careful review of available studies, it was found that there are few studies where LSSVM was adopted as the algorithm of choice for predicting the compressive strength of normal concrete. Xue [40] explored five different models and proposed a hybrid LSSVM model in one of these studies. Its hyperparameters were optimized to improve the model's estimation accuracy using improved particle swarm optimization (IMPSO) techniques. In another study, Aiyer et al. [41] explored the capability of the LSSVM and relevance vector machine (RVM) in the prediction of the compressive strength of SCC. While the RVM performed better, the margin of performance in comparison to LSSVM was minimal. Biswas et al. [35] also explored the excellent predictability of LSSVM in estimating concrete compressive strength in addition to other ML techniques. In another bid to estimate the concrete compressive strength accurately, Prayogo [42] developed a novel and advanced hybrid ML algorithm combining the robust prediction accuracy of LSSVM with a new metaheuristic, symbiotic organisms search (SOS).

The primary objective of our study is to develop a data-driven model with the ability to accurately predict the compressive strength of ternary blend (blast furnace slag, Portland cement, and fly ash) concrete using readily available experimental mixture parameters. This study proposes a least square support vector machine with coupled simulated annealing for parameter optimization (LSSVM-CSA) and genetic programming (GP) as predictive tools to solve the problem of estimating concrete compressive strength using experimentally acquired datasets. Our choice of LSSVM-CSA and GP as learning algorithms stems from their high prediction accuracy, successful deployment, and promising results recorded by many researchers [43], [44], [45], [46], [47], [48], [49], [50]. The parameters used in training the models include cement, water, coarse aggregate, fine aggregate, blast furnace slag, fly ash and superplasticizer, curing age as inputs, and compressive strength as output. In addition, the developed models were evaluated using statistical parameters such as the coefficient of determination (R²), root mean square error (RMSE), average relative deviation (ARD), and absolute average relative deviation (AARD) for the training and test datasets.