基于PI-DeepONet的储层应力场预测与泛化训练策略研究OA

In reservoir stress field simulation,Physics-Informed Neural Networks(PINN)can achieve high-precision unsupervised solutions;however,their model structure tightly couples the computational domain with physical parameters,resulting in applicability only to a fixed set of material properties.This leads to limited generalization under varying working conditions.To enhance the generalization capability of PINN,this study develops an intelligent computational approach based on the Physics-Informed Deep Operator Network(PI-DeepONet).By introducing a branch-trunk architecture and employing the Hadamard product to fuse parameter and coordinate features,an end-to-end mapping from reservoir physical parameters to the stress-displacement field is established.Furthermore,a hard-constraint mechanism and a staged progressive training strategy are integrated to construct a stress field operator model with strong generalization capacity.The results demonstrate that this method overcomes the non-generalizability of conventional PINN,achieving approximately 30.6%improvement in training efficiency under sparse physical-space discretization,and around 62.4%enhancement in prediction accuracy under dense discretization by applying hard constraints.This research provides a reliable intelligent computational framework for efficient hydrocarbon reservoir development and CO2 geological storage assessment.