Predicting Rheological Properties of Asphalt Modified with Mineral Powder: Bagging, Boosting, and Stacking vs. Single...

Area of Science:

• Materials Science
• Civil Engineering
• Computational Science

Background:

• Accurate prediction of asphalt rheological properties is crucial for pavement performance.
• Traditional single machine learning models have limitations in characterizing complex material behaviors.

Purpose of the Study:

• To systematically compare single machine learning models and ensemble methods for predicting asphalt rheological properties (complex shear modulus G* and phase angle δ).
• To develop and validate an innovative stacking ensemble model for enhanced prediction accuracy.

Main Methods:

• Fabrication of modified emulsified asphalt using emulsifiers and mineral powders.
• Rheological property testing using Dynamic Shear Rheometer (DSR).
• Data preprocessing with Local Outlier Factor (LOF), K-fold cross-validation, and Bayesian optimization.
• Development of a stacking model integrating KNN, Decision Tree, Random Forest, and XGBoost base learners with a Bayesian ridge regression meta-learner.

Main Results:

• Ensemble models significantly outperformed single machine learning models.
• The stacking model achieved the highest prediction accuracy (R² = 0.9727 for G* and R² = 0.9990 for δ).
• Shapley Additive Explanations (SHAP) identified temperature and mineral powder type as key predictive factors.

Conclusions:

• The developed stacking model provides a robust framework for optimizing asphalt mixture design.
• This approach enhances material selection and pavement performance improvement by addressing the 'black box' nature of machine learning in materials science.