Can Gaussian-Based Surrogate Models Replace Stochastic Simulations for Assessing the Undrained Stability of Spatially Variable Slopes?

Abstract

Traditional slope stability assessments often rely on deterministic methods that ignore the spatial variability of soil properties, potentially underestimating failure risk. Stochastic simulations using random field theory and advanced numerical methods such as the mixed finite limit analysis (MFLA) offer greater accuracy but are computationally demanding. This study evaluates whether Gaussian Process Regression (GPR) can be an efficient surrogate model to replace such simulations for assessing the undrained stability of spatially variable slopes. We generate thousands of stochastic simulations using MFLA with spatially random cohesion fields and train GPR models to predict the factor of safety (FoS). The surrogate's performance is evaluated using standard regression metrics, uncertainty quantification, and distributional fidelity against full simulations. Two benchmark problems-a homogeneous slope under seismic loading and an embankment dam on random clay-are used to assess generali-ability. Results show that GPR surrogates can predict FoS with high accuracy (R2 > 0.96, MAPE < 7%) and reliably capture the failure probability distribution at a fraction of the computational cost. We conclude that Gaussian-based surrogate models can effectively replace direct stochastic simulations under the tested conditions, enabling fast and probabilistically sound slope stability assessments.

Keywords: Gaussian-based surrogate models, Slope stability, Spatially variable slopes.