A GPU-accelerated adaptive active-grid MPM for high-resolution 3D modelling of large-scale landslides: Case study of the 2025 Blatten ice-rock avalanche

Large-scale landslides, such as ice-rock avalanches and debris flows, involve extreme deformation, long-distance transport, and strong interaction with complex terrain. The Material Point Method (MPM) is well suited to such processes, but its large-scale application is constrained by the memory demand of fixed background grids and the bandwidth-intensive particle-grid transfers required at every time step. This study develops a GPU-accelerated adaptive active-grid MPM framework with critical-state μ ( I ) rheology for high-resolution three-dimensional (3D) landslide simulation. An adaptive active-grid reconstruction strategy is introduced to activate only the grid blocks surrounding the evolving material domain, thereby reducing the memory-capacity demand imposed by large topographic domains. A shared memory particle-to-grid scheme is further implemented to reduce global-memory traffic, atomic-write contention, and bandwidth consumption. The framework is applied to study the 2025 Blatten ice-rock avalanche in Switzerland, which involved approximately 9.3 × 10 6 m 3 of ice and rock material. The simulation captures the main runout path, velocity evolution, and final deposition footprint. Performance analysis reveals that, at 4 m resolution, the estimated memory requirement is reduced from approximately 781 GB for a full-domain static grid to 10.8 GB, while the shared memory P2G reduction decreases the single-step computational time by 44.59%. The proposed framework enables high-resolution 3D MPM simulations on a single consumer-grade GPU, providing an efficient tool for assessment of large-scale geohazards.