Probabilistic evaluation of 3D slope stability considering soil spatial variability using convolutional neural networks

Although slope failures are inherently three-dimensional (3D) and soils exhibit spatial variability, studies that consider both factors remain limited. To address this issue, an efficient framework is proposed in this work for probabilistic evaluation of 3D slope stability with spatially variable soil properties. The framework couples a 3D convolutional neural network (CNN) with Monte Carlo simulation (MCS) for reliability assessment. Spatial variability in soil strength is modeled using random fields, with realizations obtained through the fast Fourier transform method. A deterministic solver based on discretized limit analysis (DLA) is implemented to evaluate 3D slope stability in spatially heterogeneous soils. A limited number of random field samples is generated, and the associated slope stability responses are evaluated with the DLA-based deterministic solver. The resulting paired input-output data constitute the training set for a 3D CNN, which can then estimate slope stability for new random field realizations. The trained CNN surrogate enables large-scale MCS for probabilistic slope stability analysis with high computational efficiency. Model performance is assessed at both the deterministic solver level and the CNN level. Overall, the proposed DLA-CNN coupling provides an efficient approach for 3D slope reliability analysis, enabling accurate probabilistic evaluation with markedly reduced computational cost