New Observations on the Application of LS-SVM in Slope System Reliability Analysis

Abstract

The stability evaluation of earth slopes is a common practice in geotechnical designs. To account for uncertain characteristics of soil properties, probabilistic evaluation that requires repeated calculations of factor of safety (FoS) is inevitably encountered. Because FoS for most slopes in practice is computed numerically due to the lack of analytical solutions, various surrogate models are usually developed to ease the probabilistic evaluation. This paper investigates the probability of slope system failure with surrogate models based on the least-squares support vector machine (LS-SVM) regression. First, some limitations in the current application of LS-SVM to complex slopes with multiple failure modes are pointed out. In the context of Monte Carlo simulations (MCS) for probabilistic slope stability evaluation, the authors first discuss the importance of space filling of training data to the success of the LS-SVM and then propose an efficient routine for generating the training data by which the global prediction of FoS is reasonably guaranteed. The application of the LS-SVM is illustrated through two well-documented slope examples. Comparative studies are conducted to identify the effect of training data size and hyperparameters on the model performance. It is observed from this study that the LS-SVM model can reasonably capture the global characteristics of complex slopes only when all the relevant soil layers are treated probabilistically; otherwise, some local inconsistency could be encountered. Focusing on the probability of failure prediction defined by different FoS thresholds, it is shown that the LS-SVM is robust and a promising method for the evaluation of complex slopes.