Prediction of flexural and split tensile strength of waste glass-concrete composite using machine learning algorithms

Abstract

Waste material, including glass, presents significant environmental challenges due to its non-biodegradable nature and low global recycling rates. Incorporating waste glass into concrete offers a sustainable solution, but predicting its effects on mechanical properties, particularly flexural ($f_b$) and split tensile ($f_t$) strengths, remains complex. This study utilizes machine learning (ML) algorithms (decision tree (DT), extreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), light gradient boosting machine (LightGBM), support vector regression (SVR), and gaussian process (GP)) to predict $f_b$ and $f_t$ strengths based on compressive strength ($f_c$), concrete age, and glass replacement percentage of glass-concrete composites. Thirteen experimental studies were utilized using secondary data. Results demonstrate that Pearson correlation analysis reveals strong interdependence among mechanical properties ($f_c$-f${}_b$: 0.809-0.876, $f_c$-f${}_t$: 0.927-0.948, $f_b$-f${}_t$: 0.943-0.970), with negligible influence of glass type and moderate positive impact of replacement percentage. The ML algorithms each offer unique predictive strengths—most notably, XGBoost training model achieves near-perfect accuracy (with R² equal to 0.9991). However, k-fold cross-validation revealed overfitting concerns limiting applicability to conventional concrete compositions. Non-parametric analyses reveal moderate $f_c$-$f_b$ correlations (Spearman’s $\rho =0.5879$, p=0.0739) and statistically significant $f_c$-f${}_t$ relationships ($\rho =0.6364$, p=0.0479), while ML models achieve high predictive accuracy by exploiting multi-feature interactions beyond simple pairwise correlations. These ML models enable optimized mix designs, advancing sustainable construction through efficient waste glass utilization as a partial aggregate replacement.