Random particle finite element method for probabilistic slope large-deformation analysis accounting for spatial variability

Risk assessment of slope failure requires evaluating uncertainties associated with the post-failure characteristics. This study proposes the Random Particle Finite Element Method (RPFEM) for predicting the entire process of slope failure, considering soil spatial variability. To mitigate the high computational cost of stochastic simulations, a surrogate model based on the Extreme Gradient Boosting (XGBoost) is developed. Additionally, the Sliced Inverse Regression (SIR) method is applied to reduce the dimensionality of the surrogate model associated with random field parameters. The proposed framework is validated through a classic clay slope example, wherein the surrogate model achieves high accuracy. In comparison, without employing the dimensionality reduction technique, the model performs poorly with a coefficient of determination of only 0.22 on the testing dataset. Then, a Latin hypercube sampling with 10,000 random field samples is used for probabilistic slope large-deformation analysis to assess the failure probability, runout distances, and exceedance probabilities at various distances, which can be completed in less than one second with the proposed framework. This framework offers a versatile method for the risk assessment and mitigation of landslides.