A Physics-Informed Machine Learning Model for Displacement Forecasting of Deep-Seated Landslides

Displacement-based early warning of deep-seated landslides requires displacement forecasts that are not only accurate but also physically consistent under rapidly changing hydrological conditions. Variations in rainfall infiltration, pore-water pressure, and subsurface moisture dynamics can modify the effective stress state and creep rates, leading to complex and often nonlinear displacement responses. Despite extensive modeling efforts, reliable and physically interpretable displacement forecasting remains challenging. Data-driven models often lack process consistency, while physics-based approaches are limited by uncertain parameters and simplified assumptions when applied to real-world conditions. In this study, we develop a physics-informed machine learning model that integrates physical process constraints with landslide monitoring data. A hydrologically driven deformation relationship dominated by seepage-related effects is incorporated as a model constraint to guide the prediction of displacement. The model is trained using cumulative displacement observations and hydrological forcing from an IoT-enabled in situ monitoring system deployed on a landslide, and is subsequently applied to forecast displacement over unseen periods. Results show that embedding physical constraints improves the temporal generalization and physical plausibility of predicted displacement trajectories, particularly during hydrologically triggered acceleration phases. The inferred model parameters exhibit physically interpretable and internally consistent behavior, indicating that dominant hydrological controls on deformation are captured. This framework improves both robust displacement forecasting and physical interpretability, thereby supporting the development of operational landslide early-warning systems.