Stochastic simulation of post-failure landslides in spatially variable soils is essential but computationally prohibitive with high-fidelity methods like MPM. This study introduces a Random Graph Network Simulator (RGNS) that learns underlying physics of granular flow by capturing local interactions among material points through graph neural networks, enabling accurate and efficient emulation of MPM dynamics. 

Trained on limited MPM data with Gaussian random fields, RGNS generalizes well across varying slope geometries and heterogeneity conditions. It reproduces key probabilistic metrics (e.g., runout distance, sliding volume) with high accuracy (R² > 0.93) while drastically reducing computational cost—10,000 Monte Carlo simulations in 3–4 days versus over 400 days with MPM. Results also show that deterministic approaches can underestimate low-probability, high-impact hazards. 

Overall, RGNS provides an efficient, physics-informed framework for probabilistic landslide hazard assessment.

Probabilistic landslide hazard zoning based on exceedance probability of post-failure distance.