An accelerated meshfree computational framework with machine learning classification for multi-phase modeling of landslide

This study presents a novel multi-phase computational framework that integrates physics-based modeling with machine learning algorithms for comprehensive landslide analysis from failure initiation to runout prediction. The framework employs a hydro-mechanical coupled semi-Lagrangian Reproducing Kernel Particle Method (RKPM) to model large deformation and strain localization in the seepage-induced slope failure processes. Two deformation measures, the angle change Δθ between initially orthogonal material directions and the first principal stretch λ₁, are extracted to characterize the deformation state at each computational node. K-means clustering in the (Δθ,λ₁) feature space is employed to identify shear-band localization regions, while Support Vector Machine (SVM) algorithm is introduced to identify the failure surfaces. The discrete interface points are subsequently regularized using the Idealized Curved Surface (ICS) method to generate a smooth three-dimensional failure surface suitable for post-failure debris flow simulation using a GPU-accelerated two-phase solver, the Modeling on Shallow flows with Efficient Simulation for Two-Phase Debris Flows (MoSES_2PDF). Validation against experimental data in the seepage-induced levee failure modeling confirmed the framework's high accuracy in capturing slope failure depths and runout patterns, with errors in the deformed geometry remaining below 4%. The field case study further demonstrated a scalable digital twin solution capable of delivering comprehensive landslide risk assessments.