Powering Three-dimensional Heterogeneous Slope Failure Predictions and Risk Assessment with Deep Learning

Spatially variable soil properties are routinely represented as random fields in the numerical modeling of geosystem stability. The stochastic nature of these problems requires repeated physics-based simulations for probabilistic assessment, incurring significant computational demands in large-scale three-dimensional (3D) analyses. A deep learning surrogate framework is proposed to predict the 3D spatial distribution of total displacement magnitude and the factor of safety from random field inputs. Gaussian Mixture Model clustering converts predicted displacement fields into slope failure patterns. A per-realization risk indicator is formulated by combining a continuous sigmoid-based weight with the weak zone percentage for each Monte Carlo realization, replacing the conventional binary failure criterion and yielding a distribution of risk estimates rather than a single failure probability. The approach is demonstrated on a 3D slope with spatially variable soil cohesion and friction angle. Trained with limited data, the surrogate models achieve high prediction accuracy and maintain comparable performance without retraining for random fields with vertical or horizontal scales of fluctuation different from those used in training. These capabilities provide a practical framework for large-scale 3D slope reliability analysis and comparative assessment of potential failure consequences under spatial variability, complementing the information provided by failure probability alone.