Prediction of asphalt concrete flexibility index and rut depth utilising deep learning and Monte Carlo Dropout simulation

Abstract

ABSTRACT Asphalt concrete (AC) balanced mix design (BMD) relies on laboratory testing to meet both the Illinois Flexibility Index Test (I-FIT) and the Hamburg Wheel Tracking Test (HWTT) by varying aggregate gradations and asphalt cement contents. AC mix designers would benefit from estimating the impact of constituent materials’ properties on I-FIT and HWTT before conducting a BMD trial. This study focused on the development of two deep learning models to predict I-FIT flexibility index (FI) and HWTT rut depth. The models were created based on a I-FIT database of 19,138 datasets and a HWTT database of 7602 datasets (after data preprocessing was conducted). Two deep neural networks (DNNs) were then trained to predict FI and rut depth. Monte Carlo Dropout simulations were then used in the DNN models to compute a distribution of predicted FI and rut depth. The distribution of predicted FI and rut depth provides a best estimate and range of FI and rut depth. The developed models provide a distribution of predictions with a coefficient of variation (CoV) lower than 30% for both the I-FIT and HWTT models, respectively.