Physics-informed neural networks incorporating slope stability mechanics for real-time landslide early warning in Three Gorges Reservoir Area

Abstract

Landslides pose significant threats to infrastructure and public safety in reservoir and mountainous regions, requiring accurate and physically reliable early-warning systems. This study proposes a Physics-Informed Neural Network integrated with the Marine Predators Algorithm (PINN + MPA) for real-time landslide deformation prediction and early warning in the Three Gorges Reservoir Area (TGRA). The framework integrates publicly available InSAR time-series deformation data, DEM-derived terrain attributes, and geotechnical parameters through temporal sequence construction and physics-informed learning. Slope-stability constraints are embedded into the learning process, while MPA optimizes network hyperparameters to improve convergence and prediction performance. The model achieves an MAE of 0.0246, MSE of 0.0011, RMSE of 0.0328, R2 of 0.9836, physics residual error of 0.051, and an average inference time of 3.1 ms. Compared with conventional NN, LSTM, and standard PINN models, the framework reduces prediction error while maintaining strong physical consistency and computational efficiency. The results demonstrate its capability to provide accurate deformation prediction and slope-stability assessment with minimal computational delay, supporting timely hazard identification and early-warning decisions for reservoir managers, geotechnical engineers, and disaster-management agencies.