A Physics-Informed Neural Network for Hydraulic Diffusivity Inversion in Rainfall-Induced Landslide Analysis

Rainfall-induced landslides pose a significant threat to infrastructure and communities worldwide. Their occurrence is strongly linked to the hydrodynamic behavior of slopes during intense precipitation events. Understanding the complex interactions between rainfall infiltration, pore water pressure buildup, and slope stability is critical for forecasting and mitigating these hazards. A key parameter in rainfall-induced landslide analysis is hydraulic diffusivity, which governs the rate of water movement within slopes and directly influences slope stability. However, determining hydraulic diffusivity is usually challenging because of its small magnitude, which typically ranges from the order of 10⁻⁸ to 10⁻³ m²/s. Conducting field tests or collecting soil samples from potential landslide sites for laboratory tests is time-consuming, costly, and potentially hazardous due to the challenging or unsafe accessibility of such areas. Nevertheless, the estimation of hydraulic diffusivity is essential for understanding slope hydrodynamics, predicting pore pressure buildup, and assessing the timing and severity of rainfall-induced landslides. This study presents a physics-informed neural network (PINN) inversion method to estimate hydraulic diffusivity by solving the Richards equation using pressure head data that are available from field monitoring. The proposed method focuses on treating hydraulic diffusivity as a training parameter within an inverse problem, leveraging one of the key advantages of PINNs in inversion analysis. The inverted parameter can then be utilized to assess hydrological responses induced by future rainfall events. The results of a case study show that this method can achieve an accuracy of 98.8% in estimating hydraulic diffusivity using only 30 sets of observation data with 6 designed sensors. These results highlight the proposed method’s potential to rapidly inform slope stability assessments, improve landslide forecasting, and enhance disaster preparedness in landslide-prone regions.