Application of machine learning models for prediction of rheological properties of wax-modified asphalt binders

Abstract

Machine learning models have received increasing attention in recent years to predict the mechanical behavior of different materials used in pavement construction due to their ability to configure their parameters based on data obtained in the laboratory. This study aims to develop, calibrate, and assess the performance of three multi-predictor machine learning models (Knearest neighbors, decision trees, and bagging systems) to predict four rheological variables (complex viscosity, complex modulus, phase angle, and dynamic viscosity) for two types of asphalt binders (an asphalt specified by penetration and an asphalt rubber) modified with different Carnauba wax contents (0, 3, 5, and 10% by weight of asphalt binder). Laboratory tests were carried out using a dynamic shear rheometer within temperatures and frequencies of 10 to 180 °C and 0.1 to 10 Hz, respectively. The results show that the three models have average coefficients of determination (R2) greater than 0.986 and relative errors (RE) less than 4% for predicting the rheological variables within the established temperature and frequency ranges. In addition, the results suggest the feasibility of implementing any of the three models for the multi-prediction (a single calibrated model is used to predict four variables simultaneously) of rheological variables for different modified asphalt binders. Based on these encouraging results, future applications of the models can be envisioned to predict the thermodynamic and physical properties of asphalt binders and develop multi-predictors with the ability to predict the response of multiple asphalt binders and rheological binder behavior with various additives.