Hybrid AI-Physics Approach for Thermo-Hydro-Mechanical Coupled Processes in Geological Systems

Accurate prediction of coupled thermo-hydro-mechanical (THM) processes is critical for geological applications including geothermal energy extraction, CO₂ sequestration, and high-level radioactive waste disposal. Traditional numerical simulators, while physically meaningful, are computationally expensive and time-consuming for real-time predictions and uncertainty quantification. This study presents a comprehensive comparison of three machine learning approaches for THM modeling: (1) pure data-driven Neural Networks (Pure NN), (2) Physics-Informed Neural Networks (PINN), and (3) Hybrid AI-Physics approach that combines analytical solutions with neural network corrections. Using the AnSichT benchmark test case for borehole disposal in clay formations, we trained and evaluated all three models against high-fidelity OpenGeoSys (OGS) simulations. Results demonstrate that the Hybrid AI-Physics approach achieves superior performance with R² scores of 0.9990 for pressure, 0.9996 for temperature, and 0.9950 for displacement. The Hybrid AI-Physics approach reduces mean absolute errors to 3.00×10⁴ Pa for pressure, 0.36°C for temperature, and 7.71×10⁻⁶ m for displacement, while requiring only ~ 20 minutes of training time compared to ~ 2.2 hours for PINN. The computational efficiency (over 1 million× faster inference than traditional simulators, with training overhead negligible compared to parameter definition time) and robust behavior (R²>0.9880 for all output variables across various training-test ratios) make these approaches particularly suitable for real-time monitoring, inverse problems, and uncertainty quantification in subsurface energy and geological system applications. Our findings suggest that Hybrid AI-Physics approach offers an optimal balance between physical consistency, computational efficiency, and data-driven flexibility for complex multi-physics problems in geological systems.