Integrating flume testing and machine learning for predicting rainfall-triggered slope movements

Rainfall-induced slope failures pose a growing threat to infrastructure, especially in clay-rich embankments. Previous studies have advanced understanding through field monitoring, numerical modelling, and physical testing; however, many approaches lack generalisability and often fail to capture the coupled hydraulic–mechanical interactions in unsaturated soils under varying rainfall conditions. Predictive frameworks also tend to rely on large or site-specific datasets, limiting their practical use for rapid assessment. This study presents a proof-of-concept experimental–predictive framework to evaluate the hydromechanical response of compacted unsaturated slopes under varying rainfall intensities. Using small-scale flume tests on sand–clay mixtures, rainfall intensities ranging from 20 to 120 mm/h were applied over a 120-min period to investigate time-dependent changes in volumetric water content and crest settlement. Results revealed critical thresholds between 85 and 90 mm/h for the test material, beyond which infiltration accelerated and deformation intensified. Seven regression-based machine learning models, namely, support vector regression, K-nearest neighbours, random forest, gradient boosting, decision tree, and polynomial regression (degrees two and three), were employed to forecast crest settlement based on rainfall intensity and soil water content. These models demonstrated strong predictive capabilities (R2 up to 0.987), effectively capturing the coupled hydraulic–mechanical behaviour of unsaturated slopes. The study contributes to ongoing efforts by quantifying rainfall thresholds and demonstrating the potential of interpretable, data-driven models for practical applications.