Use of an Artificial Neural Network Approach for the Prediction of Resilient Modulus for Unbound Granular Material

Abstract

The resilient modulus (M_R) is a fundamental material property that has a direct effect on the design and analysis of pavement structures. Many regression models have been developed previously to predict the coefficients of the M_R model from physical properties of base materials. However, the predicted model coefficients are confined to either a limited number of base materials or result in poor accuracy. To overcome this issue, a moisture- and stress-dependent model is adopted in this study to precisely estimate M_R of unbound base materials in unsaturated conditions, and a set of artificial neural network (ANN) models is developed to predict the coefficients of this model from base physical properties. The developed ANN models consist of seven input variables, ten hidden neurons, and one output variable. A large unbound base dataset was collected from the Long Term Pavement Performance (LTPP) database and used to train and generalize the network. Soil physical properties such as gradation (percent passing No. 3/8 sieve, percent passing No. 200 sieve), gradation shape parameter and scale parameter, index properties (i.e., plastic limit and plasticity index), maximum dry density, optimum moisture content, and test moisture content were selected as inputs for the ANN model. The M_R values estimated using the predicted coefficients were compared with the experimental data collected from LTPP and showed an R² value above 0.9, which is much higher than the M_R values computed using regression models. Finally, the M_R test results from different sources were used to validate the developed ANN models.