ANN-based dynamic modulus models of asphalt mixtures with similar input variables as Hirsch and Witczak models

Abstract

ABSTRACT Artificial neural networks (ANNs) and Gb*-based regression models were used for the prediction of the dynamic modulus (|E*|) of South Carolina’s hot mix asphalt mixtures (HMAs) the majority of which contained recycled asphalt pavement (RAP). Models’ training and testing were done using a database that contained 1656 |E*| values from 93 HMA mixtures. Gb*-based models included the Hirsch, revised Hirsch, Bari-Witczak, revised Bari-Witczak, Al-Khateeb 1, Al-Khateeb 2, NCHRP 1-40D, and the simplified global models. The results showed that Gb*-based regression models had a significant bias in prediction; Coupling VMA and Gb* had the most influence on |E*|; four-layer ANNs generally had a better performance than three-layer ANNs on using Hirsch model’s related inputs; ANN 3-15-15-1 and ANN 8-15-15-1 (developed with similar input variables as the Hirsch and Witczak regression models respectively) showed very high performance of R 2 > 0.994 on testing. Therefore, ANNs could be considered to capture the influence of the binders’ rheological properties, mixture’s volumetric properties, and RAP on |E*| of HMA mixtures far better than regression-based models.