Prediction of Dynamic Modulus of Southeastern Asphalt Concrete Mixtures using Artificial Neural Networks

Abstract

As mechanistic-empirical pavement design comes to the forefront of design, the accurate characterization of asphalt concrete (AC) through dynamic modulus (|E*|) becomes increasingly more important. |E*| captures the viscoelastic nature of AC and is essential to the accurate prediction of pavement responses under varying speed and temperature conditions. Due to the expensive and specialized equipment needed to measure |E*|, predictive models have gained popularity. However, variability in the predictive capabilities of these models from study to study indicates that they may not be applicable on a global level. Thus, the use of artificial neural networks (ANN) to predict |E*| was investigated for mixtures placed at the 2006 National Center for Asphalt (NCAT) Test Track. Comparisons were drawn with the most commonly used predictive models such that an ANN was created with inputs identical to each predictive model: Witczak 1-37A, Witczak 1-40D and Hirsch. The ANNs created for comparison with the predictive models predicted measured |E*| with great success; coefficients of determination (R2) ranged between 0.96 and 0.99, a notable improvement over the prediction capabilities of the predictive models. A correlation analysis was completed on a variety of input parameters to create an optimal ANN for the 2006 Test Track mixtures. An optimal ANN was created for the 2006 Test Track mixtures using only two inputs, effective binder content (Vbe) and the product of dynamic shear modulus of the binder (|G*|) and the sine of the associated phase angle (). This investigation also tested the prediction capability of the newly developed ANN on an independent dataset by applying the ANN to the mixtures used in the 2009 Test Track cycle.