Improved estimation of dynamic modulus for hot mix asphalt using deep learning

Abstract This study developed neural network models for the estimation of dynamic modulus ( | E ∗ | ) for hot mix asphalt (HMA) mixtures from binder properties, mixture volumetrics and aggregate gradation. The data used for training the networks were extracted from a report of the National Cooperative Highway Research Program (NCHRP) 9–19 [1] . Unlike previous parallel studies, the networks presented employed a logistic activation upon the output to ensure the positiveness and for improved accuracy of the predictions. Comparative analyses were conducted between the resulting networks and conventional predictive equations as well as networks in other previous studies. A total of 7400 records of modulus measurements were involved, among which 6700 were randomly chosen for training, 200 for validation and 500 for testing purposes. Overall, the resulting networks significantly outperformed the Witczak’s predictive equations and networks in previous studies. For the viscosity based inputs, the network achieved R 2 values of 0.984 and 0.978 on the 500 testing examples at arithmetic and logarithmic scales, respectively. For the complex modulus based inputs, the network attained R 2 values of 0.961 and 0.959 respectively for the arithmetic and logarithmic scales. An analysis of the impacts of different variables on | E ∗ | prediction revealed that the variables for aggregate gradation were not as critical as the binder properties (viscosity and complex modulus) and mixture volumetrics.