| dc.description.abstract | This study evaluates the performance of four machine learning predictive models—Artificial Neural Network (ANN), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Mathematical Empirical Prediction (MEP)—in predicting two key parameters of the Marshall test: Marshall Stability (MS) and Marshall Flow (MF). The data used were obtained from laboratory tests of hot mix asphalt (AC
WC), consisting of six primary volumetric parameters and an additional feature: the effective specific gravity of aggregates (Gse). The evaluation was conducted under two scenarios: (1) without
including Gse and (2) with Gse as an additional input feature. The results showed that XGBoost and RF consistently produced the highest prediction accuracy, particularly for the MF target. In contrast, ANN
demonstrated unstable performance, which further declined when Gse was added as a feature. The MEP model produced moderate results but remained relevant for implementations emphasizing interpretability. In general, MF was easier to predict than MS, indicating more complex inter-variable relationships in the case of stability. The inclusion of Gse significantly improved model accuracy, especially in tree-based approaches such as RF and XGBoost. This study recommends the use of ensemble tree-based machine learning models like Random Forest and XGBoost for predicting asphalt mixture quality in both laboratory
and field environments and suggests further development through hybrid approaches such as ANFIS or Quantum Machine Learning to improve prediction accuracy for stability parameters. | en_US |
