Multi‐Model Machine Learning Framework for Predicting Drainage Performance of Porous Asphalt Mixtures

ABSTRACT

Porous asphalt mixtures (PAM), such as open‐graded friction course (OGFC) and porous asphalt concrete (PAC), are widely used in permeable pavement systems due to their excellent drainage capacity, noise reduction, and skid resistance. The drainage performance of PAM is influenced by macro‐scale void characteristics and meso‐scale pore structures. Traditional empirical models often struggle to capture the nonlinear, multi‐factorial relationships inherent in these systems, especially with small sample sizes typical in laboratory studies. This study introduces a comprehensive multi‐model machine learning (ML) framework to predict four critical drainage performance indicators—permeability coefficient, limit rainfall intensity, maximum runoff, and runoff duration—using 11 measurable void and pore characteristics. Nine regression algorithms, including Gradient Boosting Regression (GBR) and Lasso, were systematically evaluated on a standardized dataset of 98 PAC‐13/OGFC‐13 specimens. A stacking ensemble with a Ridge meta‐learner was employed to enhance model performance, achieving R ²  > 0.999 across all targets. Permutation importance analysis identified the connected void ratio, total void ratio, and volumetric mean pore diameter as the dominant predictors. This framework provides a robust, interpretable, and reproducible methodology for optimizing PAM mix design through data‐driven predictions of drainage performance.