Pressure Prediction in Two-Phase Geothermal Well Flow Using Physics Informed Neural Network (PINN)

Abstract To ensure optimal production from geothermal wells, it is essential to simulate fluid behavior accurately in the reservoir and wellbore. Wellbore simulation uses mathematical models to predict the pressure drop in fluid flow from the bottom hole to the wellhead. However, it is essential to understand that using numerical simulators still has limitations due to inappropriate empirical equations and unrealistic assumptions. Artificial Neural Network (ANN) can learn the correlation between input (wellbore geometry, temperature, pressure, and mass flow rate) and output variables (pressure drop) well. ANN is very promising for solving problems in the field of science that involve processes that are not fully understood or that require complex computational processes. However, the application of ANN in the science domain has limitations because it requires extensive data, the inability to produce physically consistent predictions, and the lack of generalization to scenarios outside the training data. Physics-Informed Neural Network (PINN) integrates ANN and scientific knowledge to overcome inconsistent physical parameters with training data that is not fully representative. We will discuss the physics loss function used to produce the PINN model with the best wellbore simulation. Overall, the PINN model with physics loss functions corrected by temperature can provide more accurate pressure predictions in the wellbore and fewer errors than the ANN model. The total error produced by the PINN model for five wells is MAE 2.780 and RSME 3.151. However, the PINN model that has been developed is not suitable for use in the one-phase water zone.