Bridging factor of safety prediction and strength reduction workflows using an interpretable Transformer-enhanced ensemble model

Abstract

Accurate estimation of the factor of safety (FoS) is a fundamental task in slope stability analysis. The strength reduction method (SRM) is widely used for slope stability analysis due to its robustness and physical consistency, while machine learning (ML) techniques have been increasingly adopted to improve computational efficiency. Effectively integrating data-driven FoS prediction into established SRM workflows while maintaining physical interpretability remains an important practical objective. This study proposes an SRM-Transformer stacked ensemble model (TSEM) framework that integrates ML-based FoS prediction with physics-based strength reduction analysis. A large-scale, physics-informed FoS database containing 100,000 slope cases is constructed using Latin hypercube sampling and a validated SRM numerical program. Within this framework, a TSEM predicts FoS from key geometric and geomechanical parameters, and the predicted values are incorporated into the SRM workflow to guide the selection of near-critical strength reduction levels and reduce redundant iterations. Comparative experiments indicate that the proposed TSEM outperforms seven commonly used single and ensemble learning models, achieving an R² of 0.9857 and a mean absolute error of 0.1508 on the test dataset. SHapley Additive exPlanation analysis shows that the learned relationships are consistent with geomechanical principles, identifying slope height, cohesion, internal friction angle, and unit weight as dominant controlling factors. Framework-level evaluations demonstrate that SRM-TSEM reduces total computational time by more than 47 percent and decreases the number of strength reduction steps by nearly 58 percent relative to conventional SRM analysis, while maintaining consistent displacement and failure field patterns with average normalized displacement deviations below 6 percent. The proposed framework enhances the computational efficiency of SRM without altering its physical basis and provides a scalable and physically interpretable solution for slope stability assessment in large-scale and time-sensitive engineering applications.