Physics-informed synergy regional co-seismic landslide size prediction: A novel data-driven approach for improved reliability and interpretability

Current purely data-driven approaches to landslide size prediction primarily focus on employing landslide-related landscape factors. Few studies have attempted to integrate generalized physical parameters by encompassing slope morphology and kinetic energy transformation within given slope units (SUs) for size prediction. This leads to prediction results overly relying on landscape factors, difficulty in elucidating the dynamic origins of landslide initiation, and insufficient interpretability of the results. To address this gap, this study aims to utilize a co-seismic landslide size prediction model that incorporates generalized physics-informed. Specifically, a spatialized energy-line model is developed based on the principle of kinetic energy conservation, using SUs as the basic analytical units. From this framework, generalized physical covariates are derived to quantify kinetic energy transformation (potential maximum velocity field) and potential energy conditions (slope morphology). These covariates were then integrated into a benchmark machine learning models to optimize the prediction process. We designed a series of rigorous multi-mode cross-validation procedures and comprehensive physics-informed interpretation protocol. The experimental results demonstrate that our physics-informed-guided model significantly improves prediction accuracy, reducing the mean deviation of predicted results by 50% compared to benchmark models, highlighting its superior reliability in landslide size prediction. Through interpretability routines, we identified an interesting phenomenon: the physical parameters used not only have an independent effect on landslide size but also exhibit a nonlinear interaction between velocity variability and seismic dynamics and slope morphology parameters, presenting significant threshold characteristics and synergistic amplification mechanisms that co-regulate landslide size and are consistent with the physical mechanism of landslides. This study establishes a methodological framework that couples physical processes with data-driven modeling, offering a new paradigm for future earthquake-induced hazard simulation research.