Accuracy of Predictive Models for Dynamic Modulus of Hot-Mix Asphalt

Various models have been developed over the past several decades to predict the dynamic modulus $\mid E^{*}\mid$ of hot-mix asphalt (HMA) based on regression analysis of laboratory measurements. The models most widely used in the asphalt community today are the Witczak 1999 and 2006 predictive models. Although the overall predictive accuracies for these existing models as reported by their developers are quite high, the models generally tend to overemphasize the influence of temperature and understate the influence of other mixture characteristics. Model accuracy also tends to fall off at the low and high temperature extremes. Recently, researchers at Iowa State Univ. have developed a novel approach for predicting HMA $\mid E^{*}\mid$ using an artificial neural network (ANN) methodology. This paper discusses the accuracy and robustness of the various predictive models (Witczak 1999 and 2006 and ANN-based models) for estimating the HMA $\mid E^{*}\mid$ inputs needed for the new mechanistic-empirical pavement design guide. The ANN-based $\mid E^{*}\mid$ models using the same input variables exhibit significantly better overall prediction accuracy, better local accuracy at high and low temperature extremes, less prediction bias, and better balance between temperature and mixture influences than do their regression-based counterparts. As a consequence, the ANN models as a group are better able to rank mixtures in the same order as measured $\mid E^{*}\mid$ for fixed (e.g., project-specific) environmental and design traffic conditions. The ANN models as a group also produced the best agreement between predicted rutting and alligator cracking computed using predicted versus measured $\mid E^{*}\mid$ values for a typical pavement scenario.