Rainfall-induced instability mechanism of high-altitude gravelly soil slopes and machine learning surrogate modeling

Abstract

This study investigates the distinctive instability mechanism of high-altitude gravelly-soil slopes driven by particle gradation, weak clayey characteristics and intense rainfall. Soil parameters were obtained from laboratory tests on a typical high-steep bedding gravelly-soil slope in the Parlung Zangbo Basin, southeastern Tibet. A fully coupled stress-pore-pressure finite-element model was built in ABAQUS to simulate seepage-deformation responses under six rainfall intensities and multiple durations. An infiltration threshold of 1.9 × 10−4 cm/s is determined. This threshold, a near-surface transient saturated zone, develops rapidly and impedes infiltration. Pore-water-pressure at the slope toe responds roughly 12 h earlier than at the crest. Every 10 cm increase in cumulative infiltration produces approximately 15 mm extra toe horizontal displacement (R2 = 0.93, 95% confidence interval (CI): 1.43–1.61). A 32-h deformation lag is observed post-rainfall, representing the period required for displacement to attain 90% of its stable value. Random-forest surrogate models for toe displacement and pore-water pressure are trained on 1,296 spatiotemporal finite-element samples, yielding test-set R2 > 0.96. Comparisons with Gaussian process regression, support vector regression and eXtreme Gradient Boosting (XGBoost) reveal that XGBoost achieves optimal accuracy (horizontal-displacement root mean square error (RMSE) = 0.42 mm), while random forest provides competitive performance (mean absolute percentage error (MAPE) = 7.8%) and better interpretability. A classification model detects transient saturated zones at 92.3% accuracy. This coupled numerical-simulation and machine-learning framework provides a new tool for rapid early-warning and parameter-sensitivity analysis of similar high-altitude rainfall-induced slope hazards.