Improved predictions of asphalt concretes’ dynamic modulus and phase angle using decision-tree based categorical boosting model

The most suitable parameter to summarize the viscoelastic response of asphalt concrete (AC) mixtures is the complex modulus, defined by means of its two main components: the dynamic modulus E∗ and the phase angle φ. They are frequently determined by means of expensive and time-consuming laboratory procedures that require suitable equipment and high-skilled technicians. As an alternative, machine learning models can be trained to make very accurate predictions and thus, substitute at least some of these lab tests. This study proposes an innovative Categorical Boosting (CatBoost) approach for the simultaneous prediction of both E∗ and φ. Nine different AC mixtures were prepared, and an extensive 4-point bending test (4PBT) experimental campaign was carried out under ten loading frequencies and six testing temperatures. In order to thoroughly compare the developed model with two well-established empirical equations (Witczak-Fonseca and Witczak 1–37A), the same input features were selected. Pre-processing and resampling techniques were implemented to both reduce computational effort and improve model efficiency, whereas an in-depth sensitivity analysis was also performed. The entire methodology was implemented in Python 3.8.5. Six different goodness-of-fit metrics were used to robustly evaluate the performance of the developed CatBoost model and to compare it with the results of two regression-based models and a reference state-of-the-art artificial neural network (SoA-ANN). Findings showed that both machine learning (ML) models outperformed the regression-based ones, displaying significantly better performance for all metrics used. CatBoost and SoA-ANN showed roughly comparable results, characterized by a mean coefficient of determination (R2) slightly higher than 0.98. Since goodness-of-fit metrics resulted in no marked differences between machine learning models, CatBoost approach might be preferred because of its easy implementation in Python and its high interpretability. Within the context of pavement engineering, such an advanced machine learning model could provide a useful and powerful tool for asphalt mixtures’ design applications.