Probabilistic analysis of stress effects on an unsaturated soil slope stability using convolutional neural networks and Bayesian optimisation

Probabilistic stability analysis of unsaturated soil slope with spatial variability under rainfall infiltration is computationally intensive due to highly non-linear behaviour and numerous repeated computations. In the field, unsaturated soil typically experiences specific stress states, and these stress levels can influence soil water capacity, thereby affecting slope stability. However, such stress effects have rarely been considered in previous probabilistic analyses of unsaturated soil slope stability. The relative importance of stress effects and spatial variability on slope stability remains unclear. To tackle these issues, a convolutional neural network with Bayesian optimisation (CNNB) is proposed as a surrogate algorithm. A completely decomposed tuff (CDT) slope, which is commonly observed in Hong Kong, serves as an example. Stress effects are characterised by a stress-dependent water retention model that effectively captures the influence of stress on water capacity at any given stress level. The spatially varying soil hydraulic and mechanical parameters of the slope are simulated by multivariate cross-correlated random fields. It is found that the proposed CNNB considerably enhances computational efficiency by at least 7.7 times compared to the random finite element method combined with the random limit equilibrium method (RFEM-RLEM). Meanwhile, it maintains a reliable probability of failure (p_f) assessment with a prediction error as low as 2.9 %. Ignoring stress effects underestimates p_f of the slope by up to 90 % under rainfall in Hong Kong with a 100-year return period. Stress effects have a more significant influence than spatial variability when computing the factor of safety (FOS) of the slope. Utilising deterministic analysis without stress effects as a benchmark, the difference in FOS due to stress effects is about 3.5 times that of spatial variability. Additionally, without considering spatial variability can also lead to unsafe assessments, as evidenced by a mean FOS value of 1.04 corresponding to a 22.6 % p_f, indicating a hazardous performance level.