Physics-Informed Neural Networks for Near-Wellbore Stress Field Prediction with Enhanced Generalization

In oil and gas field development, accurate prediction of the near-wellbore stress distribution is important for ensuring wellbore integrity, particularly in tight shale reservoirs where the medium can be approximated as nearly homogeneous. The large stress gradient and multi-physics coupling of the near-wellbore stress field are key factors limiting high-precision prediction. Physics-informed neural networks (PINNs) allow integration of governing physical laws into network training processes. However, conventional PINNs cannot accurately capture local stress concentration features and lack the ability to generalize across different parameter settings. This study uses the near-wellbore stress concentration problem as an example to improve the prediction accuracy of PINNs through the incorporation of additional physical constraints and modifications to the network architecture. Furthermore, the method is extended to a physics-informed Deep Operator Network (PI-DeepONet) framework with enhanced generalization capability. The results show that, after introducing the proposed constraints, the stress field exhibits strict biaxial symmetry. The angle-adaptive residual module decreases the near-wellbore stress error from about 40–45% to less than 5%. Based on this method, the near-wellbore stress field under different in situ stress combinations can be predicted instantaneously without retraining.