Modelagem do comportamento cisalhante de descontinuidades rochosas utilizando redes neurais artificiais do tipo função de base radial

Abstract

One of the main challenges in the analysis and design of geotechnical structures implemented in rock masses is the realistic estimation of the shear behavior of discontinuities. Owing to their importance, several models have been developed. However, despite being adequate, these models have limitations, mainly due to the availability of tests to validate them, as these are large-scale tests. Alternative models based on Artificial Neural Networks (ANN) of Multilayer Perceptron (MLP) type, fuzzy logic, and neuro-fuzzy systems have also been developed. However, these models also have limitations considering that MLP requires a higher computational effort compared to other types of ANN. Regarding fuzzy logic and neuro-fuzzy systems, limitations arise from the dependence on the intervals assigned to their input variables during the modeling process. In this study, the use of artificial neural networks (ANN) based on Radial Basis Functions (RBF) was proposed for the development of alternative models to predict the shear behavior of rock discontinuities with and without infill under conditions of constant normal load (CNL) and constant normal stiffness (CNS), considering their ability to adequately handle nonlinear problems with a single hidden layer and a shorter processing time. To achieve this, various neural models were developed, and the best-performing model was selected through graphical comparisons of experimental data and validations using hypothetical discontinuities. This model was obtained using a Gaussian basis function with a spread of 0.5 and a desired mean squared error of 0.0002. The input variables included external normal stiffness (, initial normal stress (0, joint roughness coefficient (JRC), uniaxial compressive strength of intact rock (, basic friction angle of the rock (, the ratio of infill thickness to asperity amplitude (t⁄a), the infill friction angle (), and shear displacement (h). The selected model includes two output variables, that is, shear stress and dilatation. This, in turn, presents an architecture of 8-182-2 and coefficients of determination greater than 0.97 for the predicted variables, with a root mean square error (RMSE) equal to or less than 0.0255 for both the training and testing data sets. Statistical performance analyses suggested an excellent fitting model, and validations and comparisons with hypothetical and experimental data indicated that the proposed model can consistently represent the influence of input variables on shear behavior. In summary, the RBF network demonstrates the ability to model the complex relationships inherent in the shear behavior of rock discontinuities and proves to be an alternative method for estimating shear stress and dilatancy quickly and economically for everyday engineering applications, as it allows the derivation of equations for the modeled phenomenon.