Application of artificial neural networks for estimating dynamic modulus of asphalt concrete

This paper presents outcomes from a research effort to develop models for estimating the dynamic modulus (|E^*|) of hot-mix asphalt (HMA) layers on long-term pavement performance test sections. The goal of the work is the development of a new, rational, and effective set of dynamic modulus |E ^*| predictive models for HMA mixtures. These predictive models use artificial neural networks (ANNs) trained with the same set of parameters used in other popular predictive equations: the modified Witczak and Hirsch models. The main advantage of using ANNs for predicting |E^*| is that an ANN can be created for different sets of variables without knowing the form of the predictive relationship a priori. The primary disadvantage of ANNs is the difficulty in predicting responses when the inputs are outside of the training database (i.e., extrapolation). To overcome this shortcoming, a large data set that covers the complete range of potential input conditions is needed. For this study, modulus values from multiple mixtures and binders were required and were assembled from existing national efforts and from data obtained at North Carolina State University. The data consisted of measured moduli from both modified and unmodified mixtures from numerous geographical locations across the United States. Prediction models were developed by using a portion of the data from these databases and then verified by using the remaining data in the databases. When these new ANN models are used, the results show that the predicted and measured |E^*| values are in close agreement.