| Title | Application of a physics-informed neural network for inverse modeling of natural-state geothermal reservoir |
|---|---|
| Authors | Kazuya Ishitsuka, Keiichi Ishizu, Norihiro Watanabe Yusuke Yamaya Anna Suzuki, Yusuke Ohta |
| Year | 2026 |
| Conference | World Geothermal Congress |
| Keywords | |
| Abstract | Calibrating geothermal simulations is a fundamental process in geothermal development, yet its computational burden and optimization efficiency still make practical application challenging. In this study, we examined modeling capability of a physics-informed neural network (PINN) approach for the inverse modeling of natural-state geothermal system (i.e., modeling of temperatures, pressures, and permeabilities) from well and magnetotelluric measurements. PINNs are an emerging technology to incorporate partial differential equations and boundary conditions in the framework of neural networks, and train the physical laws in addition to measurements. We examined the characteristics of the proposed method to synthetic wells created from a 3D numerical model for the Ogiri geothermal field, Japan. Our results showed that the PINNs outperformed purely data driven machine learning approaches such as an artificial neural network and extreme gradient boosting in terms of the prediction accuracy of the temperature and pressure fields. As for physical interpretability, the fluid velocity field derived from PINNs was more physically plausible and consistent with the reference numerical model. We further examined how training strategies of PINNs such as loss balancing and optimizer selection influence prediction, and our results demonstrated the effectiveness of a dynamic loss balancing strategy. For the improvement of permeability fields, we demonstrated that a transfer learning strategy is an effective option to improve permeability prediction accuracy and physical validity. |