A Machine Learning Approach for the Simultaneous Prediction of Dynamic Modulus and Phase Angle of Asphalt Concrete Mixtures

Abstract

Road pavements represent the backbone of every road network. Asphalt concrete (AC) mixtures are the main technological solution for road pavement construction. Their composition must be optimized to ensure adequate structural and functional performance. One of the most reliable parameters for the characterization of AC mixtures’ viscoelastic behavior is called complex modulus. Such a stiffness property is crucial in the evaluation of pavements’ mechanical performance. The complex modulus is usually described in terms of dynamic modulus and phase angle and, to be determined, long and expensive experimental campaigns must be carried out. An interesting alternative is represented by machine learning models that could provide fast and reliable predictions if properly trained on meaningful datasets. In this paper, the results of an extensive 4-point bending test laboratory investigation are thoroughly discussed and an up-to-date artificial neural network (ANN) methodology is outlined to simultaneously predict the dynamic modulus and the phase angle of nine different AC mixtures. To summarize the performance achieved by the developed model, six different metrics were evaluated. The empirical Witczak 1-37A equation, a well-established regression model, was used as a reference to compare the performance obtained by the neural modeling in terms of dynamic modulus. Machine learning predictions showed remarkable accuracy, outperforming regression-based ones with respect to all the evaluation metrics used. Both in terms of dynamic modulus and phase angle, Pearson correlation coefficients and coefficients of determination achieved by the ANN model were higher than 0.98, resulting in a powerful and reliable predictive tool.