ABSTRACT
This study introduces a stochastic approach based on Convolutional Neural Networks (CNNs) for predicting mechanical Asphalt Concrete (AC) properties with dependency on the mixture composition, temperature and loading frequency. The underlying CNN metamodels were evaluated by a comprehensive database of AC properties with a total of 7400 dynamic modulus records. The CNN approach shows an improved accuracy compared to other state of the art machine learning approaches found in literature. Stochastic CNN based metamodels were developed to take into account the uncertainty of mechanical properties resulting from arbitrarily arranged aggregates and air voids in AC. The data used for the stochastic metamodels contain a total of 3645 dynamic modulus and phase angle values. They were obtained from microscale Finite Element (FE) simulations considering a heterogenous material composition and viscoelastic material behaviour of the AC binder. The developed stochastic CNN metamodels provide highly accurate predictions for the statistical characteristics such as mean values, standard deviations and empirical distribution functions of the dynamic modulus and the phase angle.
Acknowledgments
The authors are very grateful to Matthew Witczak, professor emeritus from the University of Maryland, for his efforts providing the comprehensive dynamic modulus database and thus supporting our work. The authors are solely responsible for the content.
Disclosure statement
No potential conflict of interest was reported by the author(s).