.png){kind=icon}
.png){kind=icon}
Return
Hierarchical graph neural networks for stochastic slope stability analysis: A coarse-to-fine surrogate modeling approach
.png){kind=icon}
PRE
AI
Z
郑
R
Q
S
DOI：10.1016/j.compgeo.2026.108505.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Random finite element methods provide a rigorous framework for probabilistic slope stability analysis but require numerous computationally expensive simulations. Existing deep learning surrogate models mainly perform direct safety factor (FS) regression and rarely represent the localized failure mechanisms. This study proposes a hierarchical graph neural network (H-GNN) surrogate that directly represents the finite element mesh as a graph. A coarse-to-fine strategy progressively predicts the potential shear band and sliding arc. The resulting probability maps guide FS regression through differentiable soft-masked feature aggregation. Homoscedastic uncertainty weighting is further introduced to balance the localization and regression tasks. The framework is evaluated using an idealized single-layer slope and the multi-layered Chicago cut slope. The H-GNN consistently outperforms the CNN and GNN-ablation baselines in FS prediction, while the Case 1 limit-state evaluation confirms accurate failure identification. Repeated experiments with different training-validation splits also demonstrate improved robustness across different training-set sizes. The intermediate probability maps expose the spatial basis of the FS prediction and provide mechanism-related evidence for interpreting the surrogate response. After offline data generation and training, the model enables rapid evaluation of large Monte Carlo sample sets. These results demonstrate that coupling failure-path localization with graph-based regression provides an accurate and mechanism-informed surrogate for repeated stochastic analyses of a prescribed slope configuration.
Keywords:
Slope stability
Spatial variability
Graph neural network
Coarse-to-fine
Mechanistic interpretability
Surrogate model
Journal
IF:
6.2
Papers:
7.0K
Citations:
2.9W
Organization
Cited Papers