Desenvolvimento de um modelo neuronal com funções de base radial para previsão dos parâmetros de resistência ao cisalhamento do solo

Abstract

The prediction of soil shear strength is essential for the development of geotechnical designs, often based on limit equilibrium theories or stress-strain analyses using the Mohr-Coulomb elastoplastic model. However, determining the strength parameters through direct laboratory or field tests can be impractical in many situations, leading to the adoption of correlations based on field tests such as the cone penetration test and the standard penetration test. The effective friction angle is commonly predicted using the blow count value from the standard penetration test as an input variable. However, some studies indicate the limitation of this isolated approach, suggesting corrections based on field stress (Kulhawy and Mayne, 1990; Hatanaka and Uchida, 1996). On the other hand, the prediction of undrained cohesion is often carried out through the penetration resistance obtained from the cone penetration test, showing significant results (Rémai, 2013; Zein, 2017; Otoko et al., 2019). Nevertheless, these correlations have limitations, as they were not calibrated for the simultaneous determination of cohesion and friction angle in soils that exhibit both resistance contributions, nor for the estimation of effective cohesion. Given the generalization ability of artificial neural networks in addressing complex problems, several studies have proposed the estimation of cohesion and friction angle from variables collected in the laboratory (Das and Basudhar, 2008; Göktepe, 2008; Shooshpasha, Amiri & MolaAbasi, 2014; Braga, 2014). This study proposes the application of artificial neural networks, specifically those with radial basis functions, to simultaneously predict effective cohesion and friction angle from parameters obtained in the standard penetration test. This aims to aid the development of designs in situations where the collection of undisturbed samples is unfeasible. To achieve this, a database with 156 samples was created and used for training and testing. The results show that the simultaneous prediction of effective cohesion and friction angle demonstrated a better generalization ability compared to other artificial neural network models, also outperforming predictions obtained through simple and multiple regressions. Additionally, the proposed models showed significant improvements in the results, reducing mean errors and enhancing estimates compared to conventional techniques. These contributions underscore the research's potential by offering an innovative approach that expands geotechnical analysis possibilities and provides greater reliability for engineering project development.