Enhancing Marshall stability of asphalt concrete using a hybrid deep neural network and ensemble learning

Accurate prediction of Marshall Stability (MS) is vital for asphalt concrete mix design and performance evaluation, yet traditional laboratory methods are resource-intensive. This study proposes and evaluates hybrid machine learning models, specifically integrating a deep neural network (DNN) base learner with various ensemble techniques (Random Forest, XGBoost, LightGBM, CatBoost, AdaBoost) through stacking, to enhance the accuracy and efficiency of MS prediction. Leveraging a comprehensive dataset encompassing binder, aggregate, and volumetric properties of asphalt mixtures, five distinct hybrid models were developed using Ridge regression as the meta-learner. Model performance was rigorously assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Coefficient of Determination (R²), Mean Absolute Percentage Error (MAPE), and Coefficient of Variation of the Root Mean Square Error (CVRMSE) on both training and unseen testing datasets. Furthermore, SHapley Additive exPlanations (SHAP) analysis was employed to interpret feature importance and model predictions. Results demonstrated that the proposed hybrid stacking models generally outperformed standalone base learners. Notably, the DNN-CatBoost hybrid exhibited superior predictive performance on the test set, yielding the lowest error metrics (MAE=0.67 kN, RMSE=0.83 kN) and the highest R² (0.86). SHAP analysis identifies Bulk Specific Gravity of Aggregate (Gsb) as the predominant predictor (31.13 % influence) of Marshall Stability, followed by VMA, Pse, and Abs. The findings indicate that the hybrid DNN-CatBoost model offers a highly accurate and robust data-driven tool for predicting asphalt concrete Marshall Stability, holding significant potential for streamlining mix design and reducing laboratory testing efforts in pavement engineering.