A High-Performance Stochastic Framework for Landslide Uncertainty Analysis Using the Material Point Method and Random Field Theory

This study proposes a novel high-performance computational framework to address the computational challenges in probabilistic large-deformation landslide analysis. By integrating a GPU-accelerated material point method (MPM) solver with a parallelized covariance matrix decomposition (CMD) algorithm for decomposing symmetric matrices, the framework achieves exceptional efficiency, demonstrating speedups of up to 532× (MPM solver) and 120× (random field generation) compared to traditional serial methods. Leveraging this efficiency, extensive Monte Carlo simulations (MCSs) were conducted to quantify the effects of spatial variability in soil properties on landslide behaviors. Quantitative results indicate that runout and influence distances follow normal distributions, while sliding mass volume exhibits log-normal characteristics. Crucially, deterministic analysis was found to systematically underestimate the hazard; the probabilistic mean sliding volume significantly exceeded the deterministic value, with 73–80% of stochastic realizations producing larger failures. Furthermore, sensitivity analyses reveal that increasing the coefficient of variation (COV) and the cross-correlation coefficient (from −0.5 to 0.5) leads to a monotonic increase in both the mean and standard deviation of large-deformation metrics. These findings confirm that positive parameter correlation amplifies failure risk, providing a rigorous physics-based basis for conservative landslide hazard assessment.