Explainable machine learning modeling for predicting the hydraulic permeability of fly ash pervious concrete

An efficient urban drainage solution, sustainable pervious concrete mitigates stormwater runoff but exhibits highly non-linear hydraulic behavior based on its constituents. This study presents an explainable data-driven framework to accurately predict pervious concrete permeability. A 70:30 holdout partitioning strategy was applied to a meta-dataset of 150 experimental records to train and validate four distinct machine learning algorithms: Artificial Neural Networks (ANN), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGB). The input space comprised critical mix design parameters (aggregate-to-binder, fly ash-to-binder, and water-to-binder ratios, alongside mean aggregate size), while SHapley Additive exPlanations (SHAP) interpreted the optimal model. Among the evaluated machine learning algorithms, the XGB model demonstrated superior predictive capability for pervious concrete permeability compared to ANN, SVR, and KNN models. Post-hoc SHAP analysis successfully decoded the model’s “black box,” proving that the aggregate-to-binder ratio and aggregate size are primary drivers governing interconnected void space. In conclusion, the developed framework replaces tedious, trial-and-error laboratory iterations with highly accurate, physically consistent virtual simulations.