Enhanced flow number prediction of asphalt mixtures using stacking ensemble-based machine learning model and grey relational analysis

The flow number (FN) is used as a key indicator of the rutting susceptibility of asphalt mixtures. However, traditional testing methods for FN are costly and complex to implement. This study aimed to develop machine learning (ML) models for predicting FN using four algorithms: Multilayer Perceptron (MLP), Support Vector Machines (SVM), Random Forests (RF), and Extreme Gradient Boosting (XGB). A comprehensive experimental database, comprising 14 distinct features and 1005 instances, was utilized for model development. Grey Relational Analysis (GRA) was applied to evaluate the significance of individual features on FN and select critical features before modeling. Furthermore, the Stacking ensemble method was employed to integrate four base models, resulting in a more robust predictor. The results indicated that the stacking ensemble-based ML model outperforms individual base models, achieving enhanced prediction accuracy for FN, with a remarkable MSE of 0.0027, MAE of 0.0134, and R² of 0.9920. Compared to other models, there was approximately a 90 % reduction in both MSE and MAE for the stacking model, underscoring the effectiveness of stacking in integrating the strengths of different base models and reducing the errors of individual models. The stacking ensemble-based ML model with GRA provides a robust and adaptable approach for accurately predicting the FN of asphalt mixtures. These findings offer valuable insights for research on asphalt pavement design. [Display omitted] •Four ML models were built to predict the FN of asphalt mixtures.•Using Stacking can generate a more effective predictor by integrating four models.•A database comprising 14 features and 1005 instances was used for modeling.•Using GRA can evaluate the significance of individual features on target.•Engineering feature VFA and test conditions play a critical role in FN prediction.