The Application of Physics-Informed Neural Networks (PINNS) for Geothermal Well Thermal Recovery Post-Injection

Abstract This study explores the application of Physics-Informed Neural Networks (PINN) to simulate thermal recovery in geothermal wells following injectivity tests. The primary objective is to implement PINN to model well thermal recovery. Real well data from the literature, encompassing parameters such as circulation time, injectate temperature, heating-up surveys, and wellbore radius measurements, are gathered for analysis. Subsequently, governing equations and initial and boundary conditions for PINN modeling are established. The Ascencio and Horner analytical methods are employed to formulate boundary conditions for PINN. The application involves predicting natural state temperatures and temperature profiles across varying radii from the wellbore to the maximum radius of interest. However, it is crucial to acknowledge certain limitations. The iterative nature of PINN requires considerable computational resources, similar to numerical models. Furthermore, assuming vertical parameter independence may oversimplify reservoir behavior. In conclusion, despite limitations, PINN demonstrates the potential to aid thermal recovery simulations. Refining natural state temperature extrapolation is deemed essential, considering a tendency to overestimate values. Validating the model’s accuracy involves refining physical parameters through backpropagation. Additionally, it is imperative to establish the model’s credibility through a rigorous comparative analysis of outcomes with those obtained from established commercial software.