A slice-based strategy for representing 3D random fields of soils for efficient slope reliability analysis using CNN model

The reliability analysis of three-dimensional (3D) slopes with spatially varying soils requires an expensive computational cost. To address this challenge, a slice-based strategy is proposed to represent 3D random fields of soil properties as multi-channel images containing complete spatial information. The convolutional neural network (CNN) is then employed as the surrogate model between the resulting images and corresponding factors of slope safety. The constructed CNN models can replace the repetitive random finite element analysis, thereby reducing the computational cost in the reliability analysis of 3D slopes. A cohesion-frictional 3D slope is presented as the case study to illustrate and demonstrate the proposed method. The results show that reasonable predictions of both the factor of safety and probability of failure for 3D slopes can be achieved using the CNN models. The slice-based strategy can characterise soil spatial variability without sacrificing spatial information, facilitating the CNNs as effective surrogate models for efficient reliability analysis of 3D slopes. Meanwhile, the prediction accuracy and robustness of CNNs are significantly improved as the number of training samples increases. Moreover, the slicing direction of the slice-based strategy can influence the prediction accuracy of CNNs.