| Abstract |
Underground geothermal storage (UGS) is a reliable, weather-independent option for thermal energy storage in cold regions, where seasonal demand peaks and remote communities face limited grid access. Designing safe and efficient UGS systems requires thermo-hydraulic (TH) models that are both physically faithful and fast enough for early-stage studies. This study presents a Physics-Informed Neural Network (PINN) framework for simulating single-phase thermo-hydraulic (TH) processes in porous media. The proposed model integrates the governing partial differential equations of mass and energy conservation directly into the loss function, enabling fully physics-based training without requiring labeled data or spatial meshing. The framework captures coupled heat and fluid transport through automatic differentiation of temperature and pressure fields, ensuring smooth computation of gradients, fluxes, and residuals across the entire domain. To improve stability, Dirichlet boundary conditions were embedded analytically using a boundary-vanishing formulation, and a staged curriculum training strategy was adopted to balance thermal and hydraulic time scales. Results show excellent agreement with analytical diffusion trends in which the temperature front advanced proportionally to the theoretical penetration depth, while the pressure field maintained a stable linear gradient. The study demonstrates that a PDE-only PINN can serve as an efficient, mesh-free alternative for modeling coupled heat and fluid flow in porous formations, providing a foundation for future extensions to heterogeneous, multiphase, or geothermal systems. |