Post-failure analysis of layered slope considering strength spatial variability using GPU-accelerated random material point method

Soil within natural slopes exhibits heterogeneity characterized by stratification and spatial variability of material properties. Nevertheless, existing large-deformation analyses that account for strength spatial variability have predominantly been conducted using single-layer slope models or have focused solely on undrained conditions. Such simplifications limit the ability to capture realistic landslide hazards of layered slopes. To investigate how coupled spatial variability of cohesion (c) and internal friction angle (φ), together with stratification, jointly governs post-failure behaviors and failure modes of landslides, this study performs stochastic large-deformation analyses of two-layer slopes by integrating random field theory. A GPU-accelerated Random field Material Point Method (GPU-RMPM) framework is developed to enable efficient large-scale Monte Carlo simulations (MCSs). Parameter sensitivity analyses are conducted to examine the effects of horizontal scale of fluctuation (θh), coefficient of variation (COV), the rotation angle of anisotropy (β), and the cross-correlation coefficient (Rcφ) between c and φ on post-failure behaviors and failure modes. The results show that shear strength spatial variability gives rise to distinct local failure (LF) and global failure (GF) modes. The GF mode mobilizes a larger sliding volume than the LF mode under a comparable influence range. Moderate horizontal heterogeneity amplifies the variability of post-failure behaviors, whereas sufficiently large θh lead to homogenized behavior and reduced uncertainty. Increasing COV and Rcφ substantially enlarge both the mean values and variability of influence range and sliding volume, increasing the likelihood of extreme scenarios. The GPU-RMPM framework achieves approximately tenfold speedup over conventional CPU-based approaches, enabling practical stochastic large-deformation analyses.