Enhanced Temperature Gradient Prediction for Asphalt Layers in Full-Depth Asphalt and Composite Pavement Using Machine Learning Techniques

Abstract

Pavement temperature is often used to adjust the back-calculated properties of asphalt layers at a reference temperature based on falling weight deflectometer measurements. The objective of this study was to enhance the prediction of pavement layer temperature gradients through machine learning (ML) techniques, focusing on full-depth asphalt and composite pavements. Continuous monitoring of pavement temperature gradients revealed that pavement surface temperature exhibited more significant fluctuations and changed more quickly with air temperature changes than internal pavement layer temperatures. As the depth of asphalt layers increased, the variation of pavement temperature gradually decreased. Beyond a certain threshold depth, daily temperature fluctuations became negligible. Applying ML models with the same set of input variables as empirical models significantly enhanced temperature prediction accuracy, with the extreme gradient boosting model emerging as the most effective. Notably, variables such as the “average air temperature of the day before the testing day” and “pavement surface temperature” were identified as key factors influencing the model’s predictions. In contrast, input variables associated with the “time of the day during testing” were deemed less critical. Consequently, the effects of including or excluding additional input variables were further examined. Adding additional input variables did not necessarily enhance the model’s performance, whereas reducing the number of input variables had only a minor impact on prediction accuracy. The study found that overall, ML techniques can significantly enhance the prediction of pavement layer temperature gradients, delivering high accuracy (R² value close to 0.95) and practical implementation possibilities using only three input variables.