Accelerating stochastic simulation of post-failure landslide runout using a random graph neural network-based simulator

Abstract

Stochastic simulation of post-failure landslide behavior under spatially variable soil properties is essential for reliable hazard assessment, yet it remains challenging due to the inherent uncertainty of geomaterials and the high computational cost of high-fidelity numerical methods like the Material Point Method (MPM). This study proposes a Random Graph Network Simulator (RGNS) to accelerate probabilistic simulations of landslide post-failure processes in spatially heterogeneous soils. The RGNS leverages graph neural networks to emulate MPM-based granular flow behavior by learning local interaction laws among material points within a latent graph representation, enabling efficient, physics-consistent, and generalizable simulation of landslide runout dynamics. The RGNS is trained on a limited set of MPM simulations incorporating Gaussian random fields and is validated against multiple slope geometries and heterogeneity parameters outside the training configuration. The results demonstrate that RGNS accurately reproduces the probabilistic distributions of key post-failure metrics, including runout distance, sliding volume, and influence distance, with coefficients of determination mostly exceeding 0.93. With its computational efficiency, RGNS enables large-scale Monte Carlo (MC) simulations; specifically, 10,000 MC realizations of landslide runout are completed in approximately 3-4 days of dynamic simulation, compared to over 400 days using a conventional MPM solver. The application of RGNS to exceedance probability-based landslide hazard zoning reveals that deterministic analyses based on mean soil properties may significantly underestimate hazard extents in low-probability, high-consequence scenarios. These results demonstrate that RGNS provides a practical means for efficiently quantifying post-failure uncertainty under spatially variable soil conditions, offering a promising pathway for accelerating probabilistic landslide hazard assessment and risk-informed decision-making.