A pre-trained deep-learning surrogate model for slope stability analysis with spatial variability

Abstract

This paper presents a pre-trained deep-learning surrogate model for the slope stability problem, which can be used to accelerate the stochastic analysis of slope stability with spatial variability. One major innovation is that the model is trained with a big dataset (>12000 data) covering common soil properties, spatial variabilities, and slope shapes such that the trained model is ready to make predictions without additional training or numerical simulations required. Other two minor contributions are: (1) special treatments for the irregular and varying boundaries of slopes and (2) novel techniques that allow the use of non-uniform mesh in data acquisitions. The proposed model is accurate with a mean-absolute-percentage-error of about 6% for the testing dataset. Seven cases of unseen data are also used to verify the model performance, including cases of different soil parameters, slope angles, and even different slope surfaces (e.g., concave and convex slopes, which are not used in training). The results show that the predict slope factor of safety is high consistent with the values from finite element simulations, and so is the obtained probability density functions. But the surrogate model takes much less computational effort (several minutes compared with hours of computing) – proving the effectiveness of our model for efficient stochastic analyses.