A Methodology for Three-Dimensional Slope Reliability Assessment Considering Spatial Variability

While the significance of spatially variable soil properties in slope stability assessment is increasingly recognized, the implementation of three-dimensional (3D) probabilistic slope reliability assessment is still hindered by its excessive computational time. This paper presents a novel and practical methodology for efficient 3D slope reliability assessment in spatially variable soils. The methodology consists of three key building blocks: 3D random finite element method (RFEM), 3D convolutional neural network (3D CNN), and a data augmentation strategy. Specifically, 3D RFEM is first employed to evaluate slope stability in spatially variable soils and generate a small set of simulation results. Subsequently, the data augmentation strategy developed based on the concept of shear strength reduction in the numerical stability calculations is implemented to expand the small data set generated using 3D RFEM. In the last step, 3D CNNs are trained using the data and employed as a deep-learning-based surrogate model to replace the computationally demanding 3D RFEM for Monte Carlo simulations (MCS) and slope reliability assessment. The synergy across the three components is illustrated using a 3D slope case study with undrained shear strength as the spatially variable random parameter of interest. The results suggest that 3D CNNs outperform other surrogate models in capturing 3D spatial information. In addition, the data augmentation strategy is not only effective in facilitating 3D CNN to handle limited training data, but it also effectively addresses the issue of data imbalance, further enhancing the robustness and accuracy of the proposed methodology. Ultimately, the proposed methodology offers an effective toolbox for conducting practical 3D slope reliability assessment, and has the potential to be implemented for other engineering applications.