PINN-Based Digital Twins for Modeling Groundwater-Induced Subsurface Collapse under Low-Observability Hydro-Mechanical Conditions

Groundwater-induced subsurface collapse presents a critical geotechnical hazard in karst terrains, which poses heavy risks to global public safety and infrastructure. Despite the substantial economic impact, predicting these failures remains challenging due to sparse subsurface monitoring and the difficulty of integrating indirect, multi-modal satellite data into traditional models. To address the challenge of low observability, we present a physics-informed neural network (PINN)-based digital twin for simulating coupled hydro-mechanical processes. The framework integrates NASA GPM (IMERG) precipitation data and Sentinel-1 InSAR surface deformation measurements to constrain subsurface dynamics. Implemented in the West-Central Florida Karst Belt, the model represents a three-dimensional domain of unconsolidated overburden overlying a weathered limestone aquifer. Subsurface dynamics are governed by transient Darcy flow and an effective stress relationship, while progressive material weakening is captured through a continuous damage variable, d, which evolves via stress redistribution and pore-pressure diffusion. Through minimizing the residuals of these governing equations, the PINN identifies the start of collapse, defined as the point where localized damage exceeds a critical threshold. Our results indicate that the digital twin produces physically consistent fields with 25–30% lower error in pore pressure and damage predictions compared to simulations that are uncoupled. Predicted collapse initiation times, Tc, remained within 18–23% of benchmark solutions, capturing time-accelerated failure during intense recharge events. Sensitivity analysis reveals that hydraulic conductivity, K, accounts for over 63% of damage variance, highlighting the model's physical interpretability. This framework provides a scalable approach for real-time hazard assessment in data-poor karst regions globally.