Prediction of compressive strength of glass powder concrete based on artificial intelligence

Abstract

The investigation of glass powder concrete (GPC) is motivated by the imperative to alleviate the pronounced environmental impact associated with conventional concrete construction and untreated waste glass. Compressive strength (CS) is a crucial criterion for assessing concrete quality. Traditionally, determining compressive strength relies on resource-intensive compression tests. This study addresses these challenges by evaluating GPC's compressive strength using fundamental and hybrid machine learning models. The study employed four machine learning models: Backpropagation Neural Network (BPNN), Extreme Gradient Boosting (XGB), Support Vector Regression (SVR), and Random Forest Regression (RFR). Additionally, four hybrid models (SSA-BPNN, SSA-XGB, SSA-SVR, SSA-RFR) were devised using the Sparrow Search Algorithm (SSA) to forecast compressive strength. Input variables included cement content, moisture content, sand content, coarse aggregate content, glass powder (GP) content, and curing age. The study reveals that optimized models generally outperform baseline models in predicting accuracy. Specifically, SSA-XGB demonstrated the highest accuracy, with R2 = 0.9645 and RMSE = 3.0640 MPa. Feature importance analysis highlights cement content and curing age as the most influential variables affecting GPC. Two-dimensional partial dependence plots (PDP) elucidate the interrelationship between GP, curing age, and cement content, confirming that, due to the volcanic ash effect and increased alkalinity, GP makes a more substantial contribution to strength. In summary, this study provides systematic recommendations for forecasting GPC, contributing significantly to both literature and practical applications. And the database containing 1045 samples will be disclosed in Appendix Afor researchers to use.