Optimizing asphalt mix design through predicting the rut depth of asphalt pavement using machine learning

Abstract

Generally, when asphalt concrete (AC) is in the design phase, the rutting development of the actual pavement is always not considered. Traditional simulative wheel-tracking tests, which are used to evaluate the rutting resistance of the designed AC, are difficult to accurately predict the rut depth of actual pavement in the realistic climate condition and traffic. To develop an asphalt mix design approach and avoid early serious pavement damage (rutting), the data relating to rut depth of AC pavements were extracted from the LTPP program,. 27 input features about climate condition, traffic, pavement structure, and pavement materials properties were selected based on collinearity diagnostics and impurity-base feature importance methods. This study also analyzed the effect of different imputation methods used to handle missing variables of the dataset on the performance of ML models. Four different ML algorithms including Support Vector Regression (SVR), Random Forest, Artificial neural networks (ANN), Gradient boosting, and multiple linear regression were trained to predict rut depth. The performances of these models were evaluated by calculating different performance measurement scoring metrics, such as coefficient of determination (R2), root means square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and scatter index (SI). Finally, an asphalt mix design procedure based on the best ML model was proposed. The results showed that compared to other imputation methods, the mean imputation method can result in the best performance of ML models. Compared with other ML models, Gradient boosting was the optimum model to predict rut depth for the imputation dataset (R2 = 0.87) and full dataset (R2 = 0.92). Additionally, each ML model except for Random Forest can achieve better accuracy than that of the calibrated rutting prediction models of HDM-4 and MEPDG from the literature. The proposed mix design procedure was implemented to design AC for the surface layer of a pavement project. The rut depth of the pavement, whose surface AC was the same as the recommended mixture design by the proposed method, is lower than the pavement failure criterion (MEPDG) after 20 years. It implicated that the proposed procedure based on the ML model can effectively avoid early intolerant rutting damage of flexible pavement.