Prediction of soil shear strength parameters using artificial neural networks

• Ano de defesa: 2019
• Autor(a) principal: Mota, Daniel Gurgel do Amaral
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Geotecnia; Redes neurais (Computação); Solos - Deformação; Artificial neural network; Soil shear strength; Cohesion; Friction angle
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/45306

Predicting soil shear strength is extremely important for the design of geotechnical designs, which often use limit equilibrium theories or stress-strain analysis with Morh-Coulomb elastoplastic model. However, obtaining the resistance parameters through direct laboratory or field tests may be unfeasible, which resulted in the wide use of correlations based on field tests such as CPT (Cone Penetration Test) and SPT (Standard Penetration Test). In predicting the effective friction angle ( '), there is a wide use of the NSPT value as the input variable, although some authors point to the inability to use this parameter only and, therefore, present corrections from the confining tension ('v0) (Kulhawy and Mayne, 1990; Hatanaka and Uchida, 1996). In the prediction of undrained cohesion (cu), the correlations point to expressive results obtained from the use of CPT penetration resistance as an input parameter in the model (Rémai, 2013; Zein, 2017; Otoko et al ., 2019). However, the use of these correlations has limited capacity, since they were not calibrated to obtain c 'and'  'simultaneously in soils with both resistance plots or even to estimate the effective cohesion (c'). Thus, considering the generalization capacity of artificial neural networks in the modeling of complex problems, some studies have been proposed to estimate c 'and ' from input variables obtained in the laboratory (Das and Basudhar, 2008; Göktepe, 2008; Shooshpasha, Amiri & MolaAbasi, 2014; Braga 2014). The aim of the present study is to propose artificial neural networks to predict c 'and ' from parameters collected in NSPT, v0 'field and soil type, in order to facilitate project design in cases where sample collection undeformed proves to be unfeasible. To this end, a database of 168 samples was collected and used for training, testing and validation of neural networks. The comparison of the models showed that the prediction of c 'and ' simultaneously had the best efficiency among the models that used RNA, also surpassing the results obtained through the linear and nonlinear correlations. The use of RNA also surpassed the efficiency of existing correlations proposed by Dunham (1954), Godoy (1983) and Hatanaka & Uchida (1996) for the prediction of  ', as well as Decourt (1989) and Terzaghi & Peck (1996). for the forecast of cu.