Explainable machine learning and experimental modeling of mechanical properties in stabilized full-depth reclaimed pavement materials

Stabilised Full-Depth Reclamation (SFDR) is a sustainable pavement rehabilitation technique that reuses in-situ materials with stabilisers. This study integrated experimental and machine learning to model key mechanical properties of SFDR mixtures (unconfined compressive strength (UCS), resilient modulus (MR), and indirect tensile strength (ITS)) as functions of ordinary Portland cement (OPC), RoadCem, cement kiln dust, fly ash geopolymer (FA/GP), asphalt emulsion, reclaimed asphalt pavement, optimum moisture content (OMC), and maximum dry density. Three hybrid gradient-boosting models were developed and automatically tuned using Bayesian Optimization. Categorical-Gradient-Boosting (CGB) achieved the highest predictive performance for UCS and MR (R2≈0.99), while Stochastic-Gradient-Boosting (SGB) delivered the most accurate ITS predictions (R2≈0.977). Extreme-Gradient-Boosting (XGB) performed strongly but ranked third. SHapley-Additive-exPlanations (SHAP) analysis identified FA/GP, OPC, and OMC as the most influential features across all predictions. A user-friendly graphical interface was developed to support rapid, cost-effective SFDR mix design in practice.