From Processing to Performance: A Multi-Scale Machine Learning Framework for SBS-Modified Asphalt Gels with Multi-Target Correlation and Explainable AI
TL;DR
A multi-scale machine learning framework is proposed to establish processing–composition–structure–performance mappings for SBS-modified asphalt gels and enables accurate and interpretable prediction of gel properties and provides a data-driven foundation for material design.
Abstract
Styrene–butadiene–styrene (SBS)-modified asphalt is a physical polymer gel system in which SBS forms a three-dimensional elastic network within the asphalt matrix. This network structure governs the rheological and mechanical properties of the material, yet the quantitative relationships among processing parameters, material composition, microstructure, and macroscopic performance remain insufficiently understood. This study proposes a multi-scale machine learning framework to establish processing–composition–structure–performance mappings for SBS-modified asphalt gels. A dataset of 1072 experimental samples was compiled from a gene database and supplementary laboratory tests. Ten input features were used to predict four performance indicators: penetration, softening point, ductility, and viscosity at 135 degrees Celsius. Four machine learning models were developed and compared. The support vector machine with radial basis function kernel achieved the highest accuracy for penetration with an R2 value of 0.9997 and for ductility with an R2 value of 0.9996. The artificial neural network performed best for softening point with an R2 of 0.9996, and extreme gradient boosting for viscosity with an R2 of 0.9993. Optuna-based optimization improved the average R2 by 2.1% over default configurations. SHAP analysis identified shear temperature, SBS dosage, and SBS particle size as the most influential factors. The framework enables accurate and interpretable prediction of gel properties and provides a data-driven foundation for material design.
