Predictive modelling and mix proportion optimisation of red mud-based geopolymer pervious concrete using machine learning

Traditional Portland cement-based pervious concrete suffers from insufficient mechanical strength and considerable environmental burdens. This study integrates laboratory testing with machine learning to predict the unconfined compressive strength (UCS) and permeability coefficient (PC) of red mud–alkali slag (RM-AS) geopolymer pervious concrete. Six machine learning algorithms were developed and evaluated, among which tree-based models demonstrated superior predictive capability. In particular, XGBoost achieved the highest accuracy, yielding a coefficient of determination (R2) of 0.89 on the test set. SHAP analysis and SEM microstructure revealed that binder composition and aggregate compactness are the dominant factors governing the mechanical and hydraulic performance. Additionally, a comprehensive scoring approach was employed, identifying an optimal mix proportion that balances strength and permeability (UCS = 19.85 MPa, PC = 5.36 mm/s), which meets the functional requirements of permeable pavement layers.