A First-Order Neural Network-Driven Method for Probabilistic Slope Stability Analysis

In this study, a practical framework is developed for intelligent probabilistic analysis and prediction of soil slope stability. A MATLAB-based program is coded to perform finite element slope stability simulations and generate synthetic datasets. These datasets are subsequently employed in an artificial neural network to identify the linear limit state function. The first-order second-moment method is then applied to perform the probabilistic analysis. Numerous numerical tests are carried out to identify the optimized deep network through Monte Carlo concepts. It is found that the network consisting of 5 hidden layers with 19, 5, 6, 57, and 64 neurons is the optimum solution in the context of the studied engineering problem. The proposed methodology is validated through two case studies. Results indicate that the neural network model is capable of predicting the mean of the safety factor with an average difference of less than 4.3% compared to conventional approved methods based on random variables and random fields. Furthermore, the findings suggest that the proposed model can rapidly and accurately estimate the performance level of a design, whereas the safety factor alone may not serve as a reliable indicator of the actual slope conditions.