Aplicação de redes neurais artificiais do tipo perceptron para previsão da condutividade hidráulica em solos não saturados

Abstract

The determination of hydraulic conductivity in unsaturated soils is essential when performing flow analysis in these porous media. However, the execution of laboratory and field tests to determine this hydraulic property is not a current practice in the scope of geotechnics, as these are time-consuming and expensive procedures. Artificial neural networks (ANN) have been widely used in soil mechanics, allowing the estimation of complex and multivariate phenomena in an easy and simple way. Thus, this dissertation aims to present a model for estimating hydraulic conductivity in unsaturated soils developed from a type of RNA known as multilayer perceptron (MLP). The model's input variables are: initial void index, initial gravimetric moisture content, percentages of sand, silt and clay, plasticity index, saturated permeability coefficient and matrix suction. During modeling, a total of 275 examples were used, of which 85% were used in the training phase, and 15% in the testing phase. The proposed model has architecture A: 8-4-2-1 and presented a correlation coefficient of 0.97 after 500 thousand iterations in both training and test phases. The results of the model adjusted satisfactorily to the experimental data used in the training and test phases, and the proposed neural network was able to represent the influence of the input variables on the hydraulic behavior of different types of soil.