Machine Learning Based Prediction of Flexural Strength of Reinforced Concrete Using Concrete Mix and Material Properties

Abstract

Accurate estimation of flexural strength is important for the assessment and proportioning of reinforced concrete, particularly when conventional concrete mixtures are modified through the incorporation of recycled aggregates and supplementary cementitious materials. The interaction among aggregate replacement, binder composition, water–cement ratio, and reinforcement characteristics makes direct prediction of flexural performance difficult using simple empirical relationships. Machine learning (ML) provides an alternative data-driven approach for capturing such interactions and identifying useful relationships between concrete composition, reinforcement parameters, and mechanical response. Previous research has demonstrated the potential of ML for predicting concrete strength and has highlighted the importance of appropriate feature selection, model validation, and interpretation when dealing with complex cementitious systems.

In this study, an ML-based framework is developed to predict the 28-day flexural strength of reinforced concrete using concrete mix and material properties. A structured dataset comprising 36 mix designs and four replicate observations for each mix, resulting in 144 observations, was developed for the analysis. The predictor variables include recycled coarse aggregate replacement, supplementary cementitious material replacement, silica fume content, water-cement ratio, steel grade, reinforcing-bar diameter, reinforcement ratio, and effective depth. The dataset was subjected to descriptive statistical analysis, Pearson correlation analysis, and variance inflation factor assessment to identify redundant predictors and control multicollinearity. Nine regression algorithms Linear Regression, Ridge, Lasso, Elastic Net, Support Vector Regression, K-Nearest Neighbors, Random Forest, Gradient Boosting, and Extra Trees were comparatively evaluated using grouped five-fold cross-validation, with observations from the same mix retained within the same validation group.

Among the evaluated models, Lasso regression produced the highest reported coefficient of determination, with an R² value of 0.984623, followed by Elastic Net (R² = 0.952104) and Linear Regression (R² = 0.912514). The corresponding Lasso model achieved a mean absolute error of 0.125826 MPa, root mean square error of 0.157385 MPa, and mean absolute percentage error of 2.124262%. The results indicate that a carefully selected combination of concrete mixture and reinforcement-related variables can provide a useful basis for data-driven estimation of flexural strength. The study also demonstrates the importance of feature screening before ML modelling, particularly when concrete mixture variables contain direct or indirect relationships. The proposed framework can support preliminary mixture assessment and reduce dependence on repeated trial-and-error evaluation, while laboratory testing remains essential for final structural and material qualification. ML-based prediction should therefore be considered a complementary engineering tool rather than a replacement for established experimental and design procedures.