Influence of metakaolin on pervious concrete strength: a machine learning approach with shapley additive explanations

This work aims to comprehensively study the application of metakaolin in pervious concrete and develop a predictive model for the properties of metakaolin blended pervious concrete. Four machine learning models - artificial neural networks (ANNs), boosted decision tress regression (BDT), support vector regression (SVR), random forest regression (RFR) and eXtreme gradient boosting (XGB) were employed to predict the influence of metakaolin on porosity, permeability, compressive strength and splitting tensile strength. The models were trained on data encompassing mix design, metakaolin content, and corresponding pervious concrete properties. Pervious investigations consistently demonstrate that adding metakaolin to pervious concrete decreases porosity and permeability. This can be ascribed to the filler effect exerted by metakaolin. Nevertheless, augmenting the substitution rate directly impacts the hydration process, resulting in a decrease in strength. The optimal substitution level of cement with metakaolin was determined to be 15-20%. Also, the results revealed that all four machine learning models accurately predicted the impact of metakaolin. ANN and XGB achieved the highest predictive power, while RFR and BDT provided reliable estimations. The sensitivity analysis reveals that the aggregate-to-binder ratio significantly influences the prediction of properties. This investigation thoroughly evaluates the properties of pervious concrete, expanding the current understanding and aiding its practical use in this field.