Physics-guided data-driven learning with active learning for slope stability analysis considering spatial variability

The spatial variability of soil properties significantly affects slope stability and failure mechanisms, introducing uncertainties in the factor of safety ( F s ), deformation responses, and critical slip surfaces (CSS). Traditional Monte Carlo-based random field methods are computationally expensive, while passive sampling-based machine learning models often suffer from inefficient data utilization. This study develops a physics-guided data-driven learning framework with active learning (PGDAL) for spatially variable slope stability analysis. The framework integrates MNMM-based high-fidelity simulations, active learning-based sample selection, and CNN/U-Net surrogate models. Spatially varying soil properties are represented as image-based inputs, while F s , displacement fields, strain fields, and CSS information obtained from MNMM simulations are used as learning targets. Four query strategies, including LCUS, margin sampling, entropy sampling, and maximum standard deviation sampling, are compared to evaluate their effects on model performance. The proposed framework enables multi-response prediction by learning the nonlinear relationship between random field inputs and slope responses, and further identifies smooth CSS through predicted deformation fields. Results demonstrate that PGDAL achieves higher prediction accuracy and faster convergence than passive sampling while requiring fewer high-fidelity samples. The proposed framework provides an efficient surrogate modeling approach for intelligent slope stability analysis considering spatial variability.