Análise comparativa da estimativa do coeficiente de permeabilidade de solos por redes neurais artificiais e métodos estatísticos

• Ano de defesa: 2023
• Autor(a) principal: Moreschi, Morgana
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Mestrado em Engenharia Civil; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Civil
• Palavras-chave em Português: Permeabilidade; Mecânica dos solos; Redes neurais artificiais; Métodos estatísticos; Solo saturado; subject.br-rjbn; Engenharia Civil
• Link de acesso: http://repositorio.ufes.br/handle/10/16829

This research analyzes and compares the performance of artificial neural networks (ANN), statistical methods and empirical and semi-empirical correlations in predicting values ​​of the permeability coefficient of saturated soils (ksat), from index properties that characterize the particle size distribution and the fine fraction. 8258 experimental data of ksat of soils composed of coarse and fine grains (2.50x10-13  ksat (m/s)  4.50x10-2 ), published in 08 databases in the literature, were compiled and compared, in order to assist in the understanding of the hydraulic properties of saturated soils and in the description of ksat prediction problems. Subsequently, a set of samples was selected and analyzed with a combination of different input variables for predicting log(ksat), using multiple linear and polynomial regression and ANN. The input variables considered were percentage of fines (silt and clay) (%Fines), liquid limit (LL), effective diameter (d10), uniformity coefficient (Cu) and void ratio (e). The results were evaluated based on the values ​​of the coefficient of determination, the root mean square error and the mean absolute error. The performance of the ANNs surpassed the regressions and correlations in the literature. Of all the results of the analyses performed, ANN302, which considered %Fines, LL, Cu and d10 as independent variables, presented, numerically, the best results. The addition of a third hidden layer reduced the accuracy of the networks. The regressions and ANNs were better than the empirical correlations for the prediction of ksat, for the investigated database, and showed that the choice of variables that characterize the particle size distribution and the fine fraction was satisfactory for the experimental database. Considering that ksat is a highly variable property and a function of several interdependent properties, the ANN technique proved to be viable, mainly because it does not require prior knowledge of the mathematical relationship between the variables and because of its ability to describe problems of greater complexity. The importance of including information on the particle size and nature of the fines in ksat databases is highlighted, mainly for characterizing the permeability of samples with hydraulic properties dominated by the fine fraction.