Comparing Artificial Neural Networks with Regression Models for Hot-Mix Asphalt Dynamic Modulus Prediction

The hot-mix asphalt (HMA) dynamic modulus ($E^{*}$) is a fundamental mechanistic property that defines the strain response of asphalt concrete mixtures as a function of loading rate and temperature. It is one of the HMA primary material inputs for common software for the mechanistic-empirical design of pavements. Laboratory testing of dynamic modulus requires expensive advanced testing equipment that is not readily available in the majority of laboratories in Middle Eastern countries, yet some of these countries are looking for implementing new pavement design methods such as those given in current standards. Thus, many research studies have been dedicated to develop predictive models for $E^{*}$. This paper aims to apply artificial neural networks (ANNs) for $E^{*}$ predictions based on the inputs of the models most widely used today, namely: Witczak NCHRP 1-37A, Witczak NCHRP 1-40D and Hirsch $E^{*}$ predictive models. A total of 25 mixes from the Kingdom of Saudi Arabia (KSA), and 25 mixes from Idaho state were combined together in one database containing $\mathrm{3,720}{E}^{*}$ measurements. The database also contains the volumetric properties and aggregate gradations for all mixes as well as the binder complex shear modulus ($G_b^{*}$), phase angle ($\delta$), and Brookfield viscosity ($\eta$). A global sensitivity analysis (GSA) was applied to investigate the most significant parameters that affect $E^{*}$ predictions. The GSA procedures based on the Fourier amplitude sensitivity test (FAST) and Sobol sequence approaches were implemented in commercially available software to evaluate the sensitivity of the three regression models to their input parameters. The ANN models, using the same input variables of the three predictive models, generally yielded more accurate $E^{*}$ predictions. Moreover, the GSA showed that aggregate, binder, and mixture representative parameters have convergent effects on $E^{*}$ predictions using one model applied, whereas binder representative parameters have the dominant effect on $E^{*}$ predictions using both of the other two models.