Development of artificial neural network predictive models for populating dynamic moduli of long-term pavement performance sections

Abstract

This paper presents a set of dynamic modulus (|E*|) predictive models to estimate the |E*| of hot-mix asphalt layers in long-term pavement performance (LTPP) test sections. These predictive models use artificial neural networks (ANNs) trained with different sets of parameters. A large national data set that covers a substantial range of potential input conditions was utilized to train and verify the ANNs. The data consist of mixture dynamic moduli measured with two test protocols: the asphalt mixture performance tester and AASHTO TP-62, under different aging conditions. The data include binder dynamic moduli values measured under different aging conditions. The ANN predictive models were trained and ranked with a common independent data set that was not used for calibrating any of the ANN models. A decision tree was developed from these rankings to prioritize the models for any available inputs. Next, the models were used to estimate the |E*| for the LTPP database materials and ultimately to characterize the master curve and shift factor function. To ensure adequate data quality, a series of quality control checks was developed and applied to grade the inputs and outputs for each prediction. Approximately 30% to 50% of all LTPP layers contained enough information to obtain reliable moduli predictions.