Investigating the effect of natural aggregate on asphalt mixture performance by integrating experimental evaluation and explainable machine learning

Abstract Natural aggregate (NA) is commonly incorporated into asphalt mixtures as part of the aggregate skeleton; however, excessive NA content may adversely affect mechanical strength and deformation resistance. This study investigates the influence of NA content on asphalt mixture performance using an integrated experimental, statistical, and explainable machine learning (XML) framework. Asphalt mixtures were prepared using crushed coarse aggregate with varying NA contents of 0%, 4%, 8%, 12%, 16%, 20%, and 24% by weight. The experimental program included Marshall stability and flow, indirect tensile strength (ITS), and wheel tracking tests to evaluate strength and deformation responses. Pearson correlation, ANOVA, and stepwise regression examined relationships between mixture variables and performance indicators. Random Forest (RF) models were then developed to predict mixture performance, achieving strong predictive capability, with testing R 2 values exceeding 0.95 and corresponding MAPE values below 2.5% for all targets. Feature importance identified NA content as the most influential variable governing mixture performance. SHAP and response surface analyses further quantified the contribution of mixture variables and illustrated their combined interaction effects. The obtained results indicated that increasing NA content reduced stability and ITS while increasing flow and rut depth. The findings also revealed that the most pronounced incremental deterioration occurred between 12% and 16% NA, where tensile strength and deformation responses exhibited high sensitivity to increasing NA replacement. This integrated framework supports reliable prediction, engineering interpretation, and informed design decisions for NA-containing asphalt mixtures.