Slope stability analysis with surrogate modeling: incorporating rainfall infiltration and spatial variability of soil parameters

Slope reliability analysis is a fundamental issue in geotechnical engineering. Traditional methods often have limitations in efficiency and accuracy when dealing with the spatial variability of soil properties and complex environmental factors such as rainfall. This paper proposes an intelligent analysis method that integrates Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and Transfer Learning (TL). To overcome the limitations of stationary random fields, non-stationary random fields are adopted to generate a more realistic representation of the spatial distribution characteristics of key soil properties (cohesion, friction angle, and saturated hydraulic conductivity). A finite element model is then constructed, and an automated workflow is implemented to perform Monte Carlo simulations, generating a dataset that accounts for inherent parameter uncertainty. The modeling process is validated against established benchmark cases to ensure reliability. To capture the time-dependent effects of rainfall infiltration, a CNN-LSTM surrogate model is developed. The CNN extracts spatial features from the soil parameter fields, while the LSTM captures the temporal variation in rainfall intensity. The integrated model jointly predicts the factor of safety (FOS), and its model performance is systematically evaluated. A transfer learning (TL) strategy is applied, and a CNN-LSTM-TL model incorporating transfer learning preserves accuracy while substantially reducing the amount of target-domain data required. This approach provides an efficient and accurate strategy for evaluating slope stability under complex environmental and spatial conditions.