Estimation of the collapse potential of loess soils in Golestan Province using neural networks and neuro-fuzzy systems

Abstract

In this paper, the calculation and estimation of the loess of samples taken from the North of Iran (Golestan Province) have been investigated. The soil used in this study has been called loess which is defined as a loose, open-structured, and metastable soil which can withstand high overburden stresses being dry, while upon saturation, the soil collapses creating enormous engineering problems. The engineering properties of the collapsible soils have been determined, which include the specific gravity, Atterberg limits, grain size distribution, and dry density. The hydrocollapsibility properties, due to wetting under different stress levels, have been measured in single-oedometer tests. Then, three neural networks have been proposed to estimate the collapse potential of soils on the basis of basic index properties. Field data, consisting of index properties and collapse potential, have been used to train and test different neural networks. Various neural network architectures and training algorithms have been examined, and a comparison study has been carried out to prove the efficiency of three types of neural networks including the multilayer perceptron (MLP) network, radial basis function (RBF) network, and adaptive neuro-fuzzy inference system (ANFIS). The effect of related parameters suppression from simulations has been analyzed. The numbers of train data and test data have been changed, and also, in-depth analysis of samples has been carried out to evaluate the efficiency of different networks. Finally, the optimal performance of estimation achieved by the best network has been presented.