A Machine Learning Approach to Facilitate Stability Analysis in Spatially Variable Soil Deposits using RS2

Slope stability analysis is a fundamental aspect of geotechnical engineering, primarily because it addresses the common issue of mobilized static shear stress in sloped grounds. Various methods are utilized in the literature to estimate the stability of infrastructures on sloped grounds, including the limit equilibrium method (LEM), finite element method (FEM), finite difference method (FDM), distinct element method (DEM), and finite element limit analysis (FELA). These numerical approaches are widely recognized in both research and practice. Each method has its own advantages and disadvantages, but more complex and versatile methods like FEM become essential for problems involving factors such as water presence, seismic loading, or the integration of structural elements like nailed walls and geogrid-reinforced slopes. The complexity further increases when considering the inherent variability of natural soil deposits. Addressing this variability involves numerous analyses through stochastic soil modeling, where the uncertainty of soil strength and stiffness parameters is accounted for. Random field theory combined with FEM can generate stability numbers for spatially variable slopes, with or without structural elements.
To accurately estimate the probability of failure or the reliability index for variable sloped grounds, important for addressing the uncertainty and risks associated with soil parameter variability, numerous numerical analyses are required. However, extensive FEM analyses are both demanding and costly, making the process less attractive. A compromise is to conduct a limited number of analyses to train a machine learning-based artificial neural network (ANN), which can then perform additional analyses within the same statistical parameter family. This article utilizes the FEM program RS2 to obtain results on the stochastic stability of a spatially variable slope. These results are integrated with a Radial Basis Function (RBF) algorithm to train and execute subsequent large-scale numerical analyses, ensuring accuracy throughout the process.

Keywords: Artificial Neural Network (ANN), Radial Basis Function (RBF), Inherent variability, Reliability-based design, Slope stability, RS2.