A rapid unsaturated infiltration prediction method for slope stability analysis considering uncertainties: Deep operator networks

Abstract

Slope stability analysis is closely related to the quantification of uncertainties in unsaturated infiltration behavior. These uncertainties arise from both internal factors, such as spatial variability in soil physical properties, and external factors, like stochastic variability in rainfall intensity. Although random-field numerical methods provide a reliable way to quantify such uncertainties, their substantial computational demands present an obvious drawback. To address this limitation, we develop a novel data-driven surrogate model using deep operator networks (DeepONet), and construct a direct mapping between the uncertain factors and their induced responses in unsaturated-slope infiltration fields. Unlike conventional deep neural networks that approximate relationships between variables, DeepONet primarily focuses on approximating relationships between functions, learning operator mappings between them (uncertain factors and their responses in slope stability). Three unsaturated infiltration scenarios under uncertain conditions are given to evaluate the performance of DeepONet. The results demonstrate the effectiveness of applying DeepONet for rapid and accurate prediction of unsaturated infiltration behavior in slopes. DeepONet exhibits an advantage in computational speed for uncertain slope stability problems, outperforming traditional numerical solvers by several orders of magnitude without requiring additional repeat training.