Machine learning-based prediction of high-strength concrete compressive strength incorporating limestone aggregates using ensemble and pruned tree models

Abstract

Accurate prediction of compressive strength is vital for ensuring the structural reliability and quality control of High-Strength Concrete (HSC). This study presents a data-driven modelling framework to predict the compressive strength of HSC incorporating varying proportions of limestone and natural aggregates, under different curing durations and ultimate loading conditions. Four tree-based machine learning models M5P, Reduced Error Pruning Tree (REP Tree), Random Tree (RT), and Random Forest (RF), were applied to a dataset comprising 123 experimental samples. The compressive strength served as the target output. Among the models, the ensemble-based Random Forest model achieved the highest prediction accuracy, with a training phase performance of CC = 0.9998, MAPE = 0.1161, RMSE = 0.2516, rRMSE = 0.23%, and NSEC = 0.9995, while testing metrics remained equally robust with CC = 0.9997, MAPE = 0.2881, RMSE = 0.3758, rRMSE = 0.38%, and NSEC = 0.9994. Sensitivity analysis using the Cosine Amplitude Method (CAM) revealed that ultimate load is the most influential input feature, with a sensitivity coefficient R_i=0.9999, indicating its dominant role in compressive strength development. Model performance was further substantiated through box plots, Taylor diagrams, and residual error visualizations. The findings support the use of Random Forest as a powerful tool for predicting the strength of HSC with blended aggregate systems, offering practical insights for performance-driven concrete design.