An accelerated meshfree computational framework with machine learning classification for multi-phase modeling of landslide
Please login to view abstract download link
We present an accelerated multi-phase landslide simulation workflow that bridges failure initiation, three-dimensional slip-surface reconstruction, and post-failure runout prediction by tightly coupling physics-based meshfree analysis with machine-learning-based interface detection. The initiation stage employs a hydro-mechanically coupled semi-Lagrangian Reproducing Kernel Particle Method (RKPM) to robustly capture extreme deformation and strain localization without remeshing \cite{Chen1996RKPM,Wei2020NSNNI}. From the RKPM solution, two deformation indicators---the angle change $\Delta\theta$ between initially orthogonal material directions and the maximum principal stretch $\lambda_1$---are extracted at particles to characterize local kinematics. K-means clustering in the $(\Delta\theta,\lambda_1)$ space provides data-driven labeling, and a Support Vector Machine (SVM) classifier then segments the evolving shear-band to identify the failure surface consistently across the domain \cite{HillmanChen2016NSNI}. The resulting discrete interface points are regularized into a smooth three-dimensional failure geometry using the Idealized Curved Surface (ICS) representation \cite{Tai2020ICS}, enabling efficient downstream simulation. For runout dynamics, we integrate the reconstructed failure surface with a GPU-accelerated two-phase shallow-flow solver (MoSES\_2PDF) to compute flow thickness, velocity fields, and hazard maps at practical runtimes \cite{Tai2019TwoPhase,Ko2021MoSES2PDF}. Benchmark and case-study results demonstrate that the proposed RKPM--ML--ICS--GPU pipeline can reproduce failure depth and runout patterns with engineering-level fidelity, while substantially reducing the end-to-end turnaround time required for scenario-based hazard assessment and digital-twin applications.
