Md Al-Amin, Md. Abdul Alim
Marshall mix design remains the predominant method for runway hot-mix asphalt pavement as outlined in the Asphalt institute manual series no-2 and the Federal Aviation Administration (FAA) general specification of the. However, the Conventional determination of Marshall mix design air voids requires multiple intricate laboratory test procedures that are time-consuming, costly, and prone to error. To address these limitations, this study developed three predictive models, such as Support Vector Regressor (SVR), Random Forest (RF), and Extreme Gradient Boosting ( XG Boost) to estimate air voids in runway pavement surface course mixes. A dataset of 155 samples was collected from the ongoing airport runway construction project at Shah Makhdum Airport, Rajshahi, Bangladesh. Five input variables were selected for modeling the air voids percentage. A Radar diagram and SHAP (Shapley Additive exPlanations) analysis were employed to visualize the feature-target relationships. Model performance was evaluated using statistical indicators including the correlation coefficient (R 2 ), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Results show that the XG Boost model achieved the highest predictive accuracy, with R 2 values of 0.829 (test) and 0.890 (train), along with lower MAE (0.196 test; 0.214 train) and RMSE (0.323 test; 0.321 train) compared to the other two models. An empirical equation was also derived, demonstrating reliable predictive capability. Moreover, a Taylor diagram was used for graphically compare model performance and identify the most effective algorithm. Overall, the developed machine learning models, particularly XG Boost is exceedingly more effective than conventional methods in terms of predictive capability for Air Voids prediction of the runway pavement mix design. By reducing dependance on extensive testing procedures, the approach provides a more accurate, time saving and cost-effective solution for airport pavement engineering.