Optimization Strategies for Critical Slip Surface Identification in Regional-Scale Limit Equilibrium Slope Stability Analyses
Please login to view abstract download link
The limit equilibrium method (LEM) has been widely utilized to evaluate slope stability. LEM aims to determine the minimum factor of safety by identifying the critical slip surface (CSS). In regional-scale analysis, CSS has traditionally been identified using grid-search procedures. To perform a CSS identification efficiently, this study introduces metaheuristic optimization methods (e.g., particle swarm optimization, PSO) to accelerate LEM. However, the performance of LEM combined with CSS optimization for regional-scale slope stability analysis has not been comprehensively investigated. To further develop landslide risk assessment, LEM with CSS optimization should be systematically evaluated using real-terrain datasets at the regional scale. In addition, sequential model-based optimization (SMBO) has not yet been applied to CSS identification, even for single-slope analyses. To make optimization methods applicable to regional-scale LEM analyses, we formulate a target-cell-based CSS identification. The slip-surface parameters for the target-cell-based CSS optimization are derived from a digital elevation model. For the slip-surface geometry, an ellipsoidal assumption is adopted to balance computational efficiency and geometric flexibility. For optimization, we employ PSO, which is known to perform well in optimizing arbitrary-shaped slip-surface in two-dimensional analysis [1], and Bayesian optimization (BO) as a representative SMBO approach. The results indicate that PSO achieves higher CSS identification accuracy than BO. In contrast, BO more effectively reduces the computational cost, particularly for iterative LEM formulations such as the simplified Bishop method [2]. Because PSO accuracy was sensitive to particle count, which is the number of candidate solutions at each iteration, improving BO accuracy under a fixed computational budget (e.g., via more sample-efficient surrogate-model updating) remains an important direction for future work.
