Neural network approach for GO-modified asphalt properties estimation

Abstract

This paper presents an innovative development process of Artificial Neural Network (ANN) to predict four properties of Graphene Oxide (GO) modified asphalt, including penetration, softening point, ductility, and viscosity. To this goal, a GO-modified asphalt database is carefully constructed and divided into 4 subsets, using input variables related to GO characteristics, mixing procedure, aging type, and properties of the initial asphalt before being modified. The model training and selection process is then conducted with random sampling techniques via Monte Carlo simulation to ensure the models’ reliability and generalizability. The results show that the selected ANN models have high performance and accuracy, with a coefficient of determination (R2) = 0.994, 0.996, 0.999, and 0.983, for penetration, softening point, ductility, and viscosity dataset, respectively. In addition, sensitivity analysis is used to evaluate the influence of input variables on the 4 properties. The findings, in good agreement with experimental results, reveal that 2 input variables, namely aging type and corresponding properties of the initial asphalt, have the most influence on the predictability of ANN models. Overall, with verified sensitivity analysis and high prediction accuracy, the proposed models could be used by material engineers to avoid costly and time-consuming experiments.