A Physics and Reliability-Informed Machine Learning Framework for Landslide Susceptibility Assessment: A Case Study in Niang-niang-pa, Gansu, China

Landslide susceptibility assessment (LSA) is essential for identifying hazard-prone regions. While machine learning (ML) has been widely used in LSA, conventional ML models often overlook underlying physical processes, limiting their robustness, especially in data-scarce or geologically complex areas. This study proposes a physics-informed machine learning (PIML) framework that integrates landslide physics and failure probabilities to enhance susceptibility predictions. The framework was applied to a 2013 rainfall-triggered landslide event in Niang-niang-pa, Gansu, China. The Simplified Transient Infiltration Model (PRL-STIM) was used to calculate the factor of safety (FoS), while the probability of failure (PoF) was derived using the first-order reliability method (FORM). These values guided ML models to ensure scientifically consistent predictions. Spatial cross-validation was employed to assess model generalizability. Results show that while the baseline ML model achieved an AUC of 0.68 on unseen regions, it exhibited significant physical inconsistencies. In contrast, the PIML model demonstrated superior physical consistency and improved generalization by 13% (AUC = 0.77). The proposed PIML framework provides a more reliable solution for LSA than either purely data-driven or purely physics-based approaches.