Prediction of large-strain cyclic behavior of clean sand using artificial neural network approach

Estimation of dynamic properties and cyclic strength becomes necessary with urbanization, for land use, planning and development. In undrained conditions, during the application of the cyclic load, the excess pore water pressure is developed in the soil, and consequently, the cyclic strength of the soil degrades and ultimately liquefaction occurs in the soil. Several factors such as site characteristics (effective confining stress, saturation condition, reconsolidation, etc.) and motion characteristics (motion amplitude and frequency) can affect the dynamic properties as well as the liquefaction phenomenon of the soil. The dynamic properties and cyclic behavior of soil can be evaluated in the laboratory using cyclic triaxial apparatus. In the laboratory, the soil behavior can be estimated either (a) by performing multiple single-stage tests or (b) by using a multistage test. Although the former procedure is more economical, the major concern is the microstructural changes during multistage testing. Therefore, as an alternative approach, a neural network-based modeling approach has been adopted in this study. The advantage of this research is that for quaternary alluvium sands with certain index properties, the proposed models can predict the cyclic behavior as well as the shear modulus of soil with significant accuracy under large strain. The database used to develop these models comprised 94 cyclic triaxial tests performed on quaternary alluvium sands with different confining pressure and loading characteristics. Results from different experimental observations are then used for validating the ANN models proposed in this study.