Accurate and Interpretable Prediction of Marshall Stability for Basalt Fiber Modified Asphalt Concrete using Ensemble Machine Learning

Abstract

Marshall Stability (MS), a parameter that reflects the load-bearing capacity and deformation resistance of asphalt concrete, is critical for pavement performance and durability. This study assesses the predictive capability of five tree-based machine learning (ML) algorithms - Decision Tree Regression, CatBoost Regressor, Random Forest Regression, Extreme Gradient Boosting Regression, Light Gradient Boosting Machine - in estimating the MS of basalt fiber - modified asphalt concrete (BFMAC). A compiled database of 128 samples was used for model training. Models were optimized with GridSearchCV and 5-fold cross-validation (CV), assessed via multiple statistical metrics, while SHAP analysis provided model interpretability. Among the tested models, Random Forest Regression (RFR) demonstrated the highest predictive accuracy (R² ≈ 0.922, RMSE ≈ 0.748 on the test set) and exhibited strong generalization capability. Interpretability analysis revealed that aggregate gradation (specifically, percentage of aggregate passing 2.36 mm and 4.75 mm sieves) and binder penetration were the most significant factors influencing MS prediction, followed by fiber content. This research underscores the potential of interpretable ML models, such as RFR, in accurately predicting MS, offering a viable alternative to conventional experimental methods for pavement material assessment.