Abstract
The study is conducted across the board from laboratory experiment to the practical use of a cutting-edge machine learning algorithm, aiming to predict the behavior of asphalt mixtures, the focus being on three aspects: Aggregate Void Percentage (AVP), Percentage of Voids Filled with Bitumen (PVFB), and Percentage of Voids in the Marshall Sample (PVMS). Approximately 200 road surface samples from Ardabil city in Iran were taken and characterized based on 11 influential features, including Optimum Bitumen Percentage (OBP), Specific Gravity of Aggregates (SGA), Asphalt and Ambient Temperatures (Aste, AmTe), Fracture Resistance, Softness, and Density. The work was complemented by employing two regression algorithms, namely XGBoost and LightGBM, on the experimental data to forecast volumetric outputs. To further improve the model's prediction stability, the ensemble (Voting and Stacking) technique was implemented, while the Artificial Protozoa Optimizer (APO) and Greylag Goose Optimization (GGO) were employed for hyperparameter tuning to achieve better generalization and higher accuracy. To overcome the curse of dimensionality and enable the interpretability of the parameters, the authors first performed feature selection followed by a sensitivity analysis. Results show that XGBoost can offer excellent values for both R(2) and RMSE for all output variables. The methods of ensemble and optimization are among the factors that further push these values. These results demonstrate the potential of combining traditional pavement experimental data with machine learning algorithms for predicting asphalt mixture performance, hence paving the way for a more cost-effective pavement design and management.