ABSTRACT
In this paper, a neural network approach is used to describe the mechanical behavior of unsaturated soils. The network has a sequential architecture, that is, a multi-layer perceptron network with feedback capability. The input layer consists of ten neurons. Four of the input neurons, namely, initial water content, dry density, degree of saturation and suction represent initial soil conditions. The remaining six neurons, namely, axial strain, volumetric strain, deviatoric strain, net mean pressures with respect to air and water pressure, and change in suction are continuously updated for each increment of axial strain based on outputs from the previous increment. The output layer consists of three neurons representing values of deviatoric stress, volumetric strain, and change in suction at the end of each increment. Next, a database was developed from triaxial test results available in literature. The database was used to train and test the network. Neural network simulations were compared with experimental results. The comparison indicates the good performance of the proposed network for predicting mechanical behavior of unsaturated soils.
