Prediction of Nonlinear Moisture Profile in Pavement Unbound Layers Using a Machine Learning-Based Method

Abstract

Unbound base and subgrade provide fundamental support for conventional flexible pavements. The moisture profile within these layers directly affects their strength and thus the pavement performance. However, it is challenging to obtain reliable moisture profiles. This study aims to predict the moisture content at various depths of pavement unbound layers using a machine learning method, with a focus on identifying the influencing factors that affect moisture content. The framework employs the Bayesian neural network algorithm to develop the prediction model, encompassing comprehensive datasets including pavement structure, material, climate, and moisture profiles, which were collected from the Long-Term Pavement Performance program. The SHAP (SHapley Additive exPlanations) approach was applied to identify the most impactful input variables. The results show that the developed predictive model demonstrates good performance in predicting the moisture content at different depths of unbound layers. The average relative humidity is found to be the most critical factor for moisture prediction, regardless of the depth. The temperature, 3/8-in. sieve passing percentage of base layer material, and No. 200 sieve passing percentage of subgrade material also play significant roles in prediction. The accurate prediction of the moisture profile is important for determining the modulus and strength of pavement foundation, which can further contribute to more informed pavement design and maintenance strategies, especially under the changing climate.