Boosting Hot Mix Asphalt Dynamic Modulus Prediction Using Statistical and Machine Learning Regression Modeling Techniques

Abstract

The prediction of asphalt mixture dynamic modulus (E*) was investigated based on 1128 E* measurements, using three regression and thirteen machine learning models. Asphalt binder properties and mixture volumetrics were characterized using the same feeding features in the NCHRP 1-37A Witczak model. However, three aggregate gradation characterization approaches were involved in both modelling techniques: the NCHRP 1-37A gradation parameters, Weibull distribution factors, and Bailey method parameters. This study evaluated the performance of these models based on various performance indicators, using both statistical and machine learning regression modeling techniques. K-fold cross-validation and learning curve analysis were conducted to assess the models’ generalization capabilities. The conclusions of this study demonstrate the superiority of the ML models, particularly the Catboost ensemble learning regression (CbR). Hyperparameter optimization and residual analysis were performed to fine-tune and confirm the heteroscedasticity of the CbR model. The Bailey-based CbR model showed the highest coefficient of determination (R2) of 0.998 and the lowest root mean square error (RMSE) of 220 MPa. Moreover, SHAP values interpreted the CbR model and showed the relative importance of its feeding features. Based on the findings of this study, the CbR model is suggested to accurately predict E* for a variety of asphalt mixtures. This information can be used to improve pavement design and construction, leading to more durable and long-lasting pavements.