Parameter Estimation of Reinforced Soil Based on Neural Networks

This paper presents the application of neural networks for estimating constitutive parameters of a nonlinear elastic model. The model is used to describe mechanical behavior of soil reinforced with fiber and lime. First, shear modulus and soil strength are assumed to be unknown nonlinear functions of multiple variables such as fiber and lime contents, confining pressure and sample aging periods. Then, a multilayer neural network is designed to map the highly nonlinear functions. Finally, conventional triaxial shearing tests are conducted with nine groups of soil samples to provide the experimental data for training and testing the neural network. The neural network model is compared to a linear regression model with exponential parameters. Results indicate the neural network approach is more efficient and the neural model provides much higher accuracy than the linear regression model.