Prediction and interpretability study of permeability coefficient in pervious concrete using machine learning and optimization algorithms

Permeability is a key property of pervious concrete, yet laboratory measurements are time-consuming and labor-intensive, and previous machine learning studies often overlooked testing-related factors. To address this limitation, a hybrid machine learning framework was developed to predict the permeability coefficient (PEC) of pervious concrete. After extracting and cleaning data from published literature, 450 valid samples were compiled. Input features included water–cement ratio, aggregate–cement ratio, aggregate size parameters, specimen geometry, porosity, and testing method, with PEC as the output. Five machine learning algorithms (support vector regression, artificial neural networks, random forest, LightGBM, and XGBoost) were combined with three optimization algorithms to construct 15 hybrid models. Among them, the PSO-optimized XGBoost model achieved the best performance, with an R² of 0.990 and RMSE of 0.790. SHAP analysis revealed that porosity had the most significant influence on PEC, followed by testing method, aggregate size, and mix proportions. Partial dependence plots further illustrated the effects of key variables. A user-friendly prediction platform was developed to enable rapid PEC estimation and SHAP-based visualization, supporting data-driven mix proportion optimization. This study enhances PEC prediction accuracy and provides practical guidance for pervious concrete design and performance improvement.