Machine learning models for predicting the compressive strength of cement-based mortar materials: Hyper tuning and optimization

This study utilized machine learning approaches, encompassing gradient boosting (GB), gene expression programming (GEP), decision trees (DT), support vector regression (SVR), random forest (RF), SVR with Bagging (SVR-Bagging), and SVR with Boosting (SVR-Boosting), to develop a robust model for estimating the compressive strength (CS) of mortar with or without metakaolin. GB algorithm emerged as the optimal choice. Notably, GB achieved the greatest coefficient of determination value (0.99), signifying its superior predicting power compared to SVR-Boosting (0.94), DT (0.92), GEP (0.91), RF (0.90), SVR-Bagging (0.87), SVR (0.79). Additionally, GB exhibited the lowest mean absolute error (MAE) score (2.07), indicating its accuracy in estimating the CS of metakaolin mortar with minimal deviation from the experimental values. Furthermore, SHapley Additive exPlanations (SHAP), individual conditional expectation plots, and partial dependence plots were utilized to provide a deeper interpretation of the model results. Age of mortar demonstrated a strong positive correlation with CS, while water-to-binder ratio and binder-to-sand ratio exhibited negative associations. Additionally, maximum diameter of aggregate influenced CS negatively, whereas metakaolin-to-binder ratio and superplasticizer content showed positive correlations. A user-friendly graphical interface was developed to predict the CS of the mortar, providing immediate results without the need for traditional, time-consuming experimental procedures.