Document Type : Research Paper
Authors
1 PhD Student, Faculty of Science, Ferdowsi University of Mashhad, Iran
2 Professor, Faculty of Science, Ferdowsi University of Mashhad, Iran
3 Associate Professor, Soil Conservation and Watershed Management Research Institute, Agricultural Research, Education and Extension Organization (AREEO), Tehran, Iran
Abstract
The swelling potential of fine-grained soils is one of effective parameters on soil mechanical behavior and erosion and fundamental data required for the design, construction and choosing construction materials. This paper presents a multi-layer perceptron (MLP) artificial neural network (ANN) model to prediction of the swelling potential of marl soils. Marl soil is a fine-grained soil. The Levenberg-Marquadt learning algorithm was used to train the networks. Existing models prediction of soil swelling potential based on physical and soil index parameters. The present study considers the effects of chemical factors on the behavior and characteristics of fine-grained soils along with the common soil index parameters. The model used physicochemical and mechanical test results from 60 marl soil samples taken from marl formations in the Neogene basin in central Iran (Tehran, Qom and Saveh regions). The models were designed to use different input data sets and structures to determine which soil properties and ANN structures correlate well with the swelling potential parameter. Electrical conductivity (EC) of saturated soil was a new input parameter used in addition to the physical and soil index parameters that include the atterberg limit, activity, content of the clay and silt, initial of porosity ratio and dry density. Values of RMSE, R2 and MCE (evaluation criteria) related to the best model with the physical parameters LL, PI, A, M, C and Yd0 are respectively 0.89, 2.3, 0.84, and for the best model with the physical parameters LL, PI, M, C, Yd0 and EC are respectively 0.92, 1.7, and 0.91.The results of the evaluation criteria models show that inclusion of EC improved the accuracy of the model. It was found that the accuracy of the generalizations and estimations of the ANN models was further increased by clustering data before the data division stage by k-means method to Compared with hierarchical method.
Keywords
)