Data-driven predictive models for evaluating optimum binder content and volumetric properties of bituminous mixtures using design variables

Abstract

This study introduces data-driven predictive models to optimize pavement design by accurately estimating the optimum binder content (OBC) and key volumetric properties of bituminous mixtures such as bulk specific gravity (Gmb), Air void (AV), Void in Minerals Aggregate (VMA) and VFB (Voids filled with bitumen). Using machine learning algorithms such as Linear Regression (LR), Decision Tree (DTR), Random Forest Regression (RFR) and Gradient Boosting Regression (GBR), the research analyzed a comprehensive dataset of 960 samples incorporating variables like aggregate source, aggregate gradation, design aggregate gradation, binder type and compaction levels. The study found out that both GBR and RFR models (R2 ≥ 0.962, RMSE ≤ 1.277, MAE ≤ 0.83 and MAPE ≤ 5.175 %) significantly outperformed LR and DTR showcasing model's suitability for accurate and reliable predictions. These results highlight their ability to handle complex nonlinear interactions between design parameters. A key innovation of this study is the integration of SHapley Additive exPlanations (SHAP) analysis which quantified the impact of design factors, identifying compaction and aggregate gradation as the most influential variables across all volumetric properties while binder type and aggregate source exhibited moderate effects. In the context of OBC, the aggregate source was the most significant variable. Moreover, a graphical user interface (GUI) was developed to facilitate the practical application. This research study sets the stage for adoption of data-driven methodologies to improve the durability and efficiency of road infrastructure. It also highlights the transformative potential of machine learning in the field of pavement engineering. © 2025 Elsevier Ltd