A machine learning-aided surrogate model for time-dependent reliability analysis of Baishuihe landslide under rainfall considering spatially variable soils

Abstract As a major landslide-prone region in China, the Three Gorges Reservoir Area (TGRA) requires rigorous reliability assessments of slope stability. However, the coupled effects of rainfall infiltration, soil spatial variability, and shear strength parameter degradation (SSPD) due to wetting–drying cycles on reservoir slope stability remain ambiguous. To bridge this gap, this study proposes a machine learning-aided surrogate modeling framework. It integrates three key components: Karhunen–Loève (K-L) expansion for representing cross-correlated non-Gaussian random fields, sliced inverse regression (SIR) for variable dimensionality reduction, and an extreme gradient boosting machine (XGBoost) optimized by the enhanced whale optimization algorithm for slope response prediction. The KL-SIR-XGBoost framework offers considerable advantages over existing commercial software combined with Monte Carlo simulations, enabling more efficient and adaptable reliability analyses for large-size unsaturated slopes in spatially variable soils under complex transient seepage and strength degradation processes. The proposed framework is applied to the Baishuihe landslide in TGRA, demonstrating high computational efficiency without sacrificing accuracy. The results quantitatively demonstrate that intense and long-term rainfall notably exacerbates the slope reliability and amplifies the landslide volume, with more pronounced effects under successive SSPD. The findings of this study highlight the critical importance of incorporating rainfall infiltration and SSPD as key factors into the reliability-based slope design or landslide risk assessment in reservoir areas.