Predicting Compressive Strength of Pozzolanic Concrete by Mixing Various Admixtures Using Analytical and Machine Learning Approaches

Data regularization and removing noise from data are vital issues in evaluating the compressive strength (CS) of pozzolanic concrete. Because several admixtures are used in pozzolanic concrete in various percentages, past studies used machine learning (ML) techniques to predict the impact of admixture percentages on CS. It is still challenging to predict CS by considering the impact of such admixtures. Various ML techniques may be capable of overcoming these challenges. For this reason, the present study performed a comprehensive comparative analysis of analytical and ML approaches to predict the CS of pozzolanic concrete. A total of 1030 published experimental data points were utilized to develop five regression-based analytical and ML models. Seven statistical metrics were used to evaluate model performance. Eight variable parameters of experimental data were used to predict CS, such as cement, blast furnace slag, fly ash, water, superplasticizer, coarse aggregate, fine aggregate, and curing age of concrete. Random Forest (RF) shows the highest accuracy in predicting CS compared to other models within the CS range of 20 to 40 MPa. The R2 value of RF is 0.99, which indicates good cross-validation. RF shows 15% higher accuracy compared to the best analytical model, although ML performs better than the analytical model. Concrete age and cement content represent the most effective factors in predicting CS based on SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs), while water exhibits a strong negative impact. Therefore, this study may be used as a guideline to predict CS for the development of sustainable concrete mix design.