Optimizing Asphalt Compaction: Intelligent Compaction Roller Frequency and Machine Learning Prediction of In-Place Density

Traditional destructive methods for assessing asphalt pavement compaction, such as core sampling, are time-consuming, costly, and provided limited spatial coverage. Nondestructive approaches are essential for real-time quality control (QC) and assurance (QA). Intelligent compaction (IC) offers continuous stiffness measurements using vibratory rollers, supporting efficient compaction monitoring. This study investigates the effect of intelligent compaction roller frequency on IC measurement values (ICMVs) and predicts in-place density using machine learning models based on ICMVs and nuclear density gauge (NDG) data. Three models—support vector regression (SVR), random forest (RF), and CatBoost—were optimized with Bayesian optimization algorithm (BOA) to predict ICMVs from vibratory roller frequency, roller speed, number of passes, layer thickness, and asphalt mat temperature. CatBoost outperformed others in predicting ICMVs [coefficient of determination (R2)=0.77, root mean square error (RMSE)=4.29, and mean absolute error (MAE)=3.14]. Shapley additive explanations (SHAP) analysis identified frequency setting as the most influential factor in energy transfer during compaction. Furthermore, CatBoost achieved an R2 of 0.72, RMSE of 1.00, and MAE of 0.80 for predicting in-place core density. SHAP beeswarm plots confirmed that higher NDG readings are strongly correlated with higher core density. Field data collected from three IC rollers operating at frequencies of 62.84, 63.74, and 64.63 Hz and speeds of 3.97 km/h (2.47 mi/h), 6.41 km/h (3.98 mi/h), and 8.36 km/h (5.19 mi/h) revealed that a 1.79-Hz increase in each roller frequency, when combined with speeds above the optimal 3.4 km/h (2.1 mi/h)−5.4 km/h (3.4 mi/h) range, resulted in a 140% reduction in Hamm measurement value (HMV), a 5.6% decrease in NDG, and a 2.7% reduction in core density. This highlights the need for site-specific optimization of IC roller frequency and speed.