Integrated finite element and artificial neural network methods for constructing asphalt concrete dynamic modulus master curve using deflection time-history data

• Main Authors: Hamim, Asmah, Md. Yusoff, Nur Izzi, Omar, Hend Ali, Jamaludin, Nor Azliana Akmal, Abdul Hassan, Norhidayah, El-Shafie, Ahmed, Ceylan, Halil
• Format: Article
• Published: Elsevier Ltd 2020
• Subjects: TA Engineering (General). Civil engineering (General)
• Online Access: http://eprints.utm.my/id/eprint/90993/; http://dx.doi.org/10.1016/j.conbuildmat.2020.119549
• Institution: Universiti Teknologi Malaysia

Dynamic modulus |E*| is one of the essential material properties input in the American Association of State Highway and Transportation Officials (AASHTO) Mechanistic-Empirical Pavement Design Guide (MEPDG). Asphalt concrete (AC) dynamic modulus master curve is used to determine the modulus of asphalt concrete over a wide range of temperature and frequency. However, the standard laboratory test procedures for establishing asphalt concrete |E*| and plotting the AC |E*| master curve are time consuming and require considerable resources. Therefore, this study aims to predict AC |E*| master curve by using data from a falling weight deflectometer (FWD) deflection time-history. Prior to developing the model, a simple performance testing (SPT) dynamic modulus test was conducted in the laboratory on five core specimens to obtain the dynamic modulus data at several test temperatures and load frequencies. Results of SPT dynamic modulus show that the |E*| of all specimens is influenced by both loading rate and test temperature. The specimens are stiffer at low temperature and high frequency, and the |E*| values are the lowest at the highest temperature and lowest frequency. Artificial neural network (ANN) models are designed using the FWD deflection-time history data obtained by the finite element method (FEM) to predict the AC |E*| master curve. This study uses two types of ANN models, namely multilayer feed-forward neural network (MLFN) and radial basis function network (RBFN). ANN results show that both MLFN and RBFN models have a promising potential in the construction of AC |E*| master curve. A comparison of the two types of ANNs revealed that RBFN has a lower percentage of error, and is therefore more accurate than MLFN.