Estimating swelling characteristics of clays using methylene blue test - a machine learning approach /


Thesis Type: Postgraduate

Institution Of The Thesis: Middle East Technical University, Turkey

Approval Date: 2014

Thesis Language: English

Student: Gamze Didem Öget

Co-Supervisor: ERDAL ÇOKÇA, ONUR PEKCAN

Abstract:

Clayey soils tend to increase volume when they interact with water, by a phenomenon known as swelling. It is a major problem worldwide causing excessive economical damage for the infrastructure that needs to be taken into consideration. In order to avoid the damage, the identification of swell susceptible soils and predicting their swelling potential is a must. Our study mainly focuses on the prediction of swell potential of clayey soils using methylene blue (MB) test. A set of laboratory tests containing the physical properties of clays, MB tests and oedometer tests are performed. For this purpose, 32 samples obtained from different regions of Turkey are tested to obtain Atterberg limits, clay contents and methylene blue values (MBVs). In addition, maximum dry density, optimum water content, swell percent and swell pressure tests are conducted on 20 of these 32 samples. Then the laboratory data with similar characteristics available in the literature are compiled and combined with our data set to generate a comprehensive data base. Using this database, the swelling potential is examined such that the swell percent and MBV are predicted through physical characteristics. First multivariate linear regression technique is used to understand the relationships in the database. However, the results show that the variables in the database are not linearly correlated. Then a machine learning approach is utilized such that Genetic Expression Programming (GEP) and Artificial Neural Networks (ANN) are applied to understand the relations. The results prove that the nonlinear relationship can best be modeled using ANNs. The values of MAPE for the best models of Dataset I, II, III, and IV for MBV prediction are 4.2%, 5.0%, 11.5%, and 30.6%, respectively. The ones for the determination of swell percent for Dataset I, and II are 1.8% and 20.7%, respectively.