Predicting the dynamic modulus of asphalt mixture using machine learning techniques: An application of multi biogeography-based programming

Abstract The dynamic modulus E ∗ of asphalt mixtures can be used to characterize the behavior of asphalt pavements at a wide range of traffic and climate conditions. The use of E ∗ predictive models instead of direct laboratory-based measurements can provide several advantages as they do not need trained personnel and expensive equipment. In this study, biogeography-based programming (BBP) was used to develop E ∗ predictive models with improved accuracy compared to previously developed models. For this purpose, two models with different architectures were developed using a dataset containing information on 4022 asphalt mixture samples. Another dataset including the records of 90 asphalt mixtures was used for testing the developed models and comparing their performance with some of the most commonly used models for the prediction of E ∗ . The results showed that both architectures provided E ∗ predictive models with excellent accuracy. Moreover, the developed models were found to outperform the Witczak model, Hirsch model, and ANN model. The first BBP model included only four variables: temperature (T), frequency (F), voids in mineral aggregate (VMA), and low-temperature PG (PGL). The second BBP model included eight variables: T, F, VMA, PGL, high temperature PG (PGH), asphalt content (AC), volume of effective bitumen content (Vbeff), and recycled asphalt pavement (RAP) content. A parametric study and a sensitivity analysis indicated that T and F were the most influential factors affecting the values of E ∗ .