High-fidelity simulation of physical fields in complex structures is critical for structural analysis and engineering design. However, the high computational cost of the Finite Element Method (FEM) restricts its real-time application. To address this bottleneck, a novel method for rapid stress field calculation (GP-GatED) is proposed based on a geometry-physics graph attention encoder-decoder network. This method maps finite mesh elements to graph nodes, fusing geometric attributes, physical features, and global boundary conditions to construct an enhanced graph structure. Based on the shared topological relationships among mesh elements, an edge weight function is introduced to characterize the mechanical coupling. To meet the efficiency demands of training large-scale mesh models, this paper introduces a subgraph construction strategy based on circumferential angle partitioning, leveraging geometric rotational symmetry. Additionally, a parallel training scheme is integrated to accelerate large-scale data processing. Validated on wind turbine main shaft bearings, experiments under 300 operating conditions show that the predicted Von-Mises stress fields highly agree with finite element benchmarks. The coefficient of determination R2 reaches 0.999 8, and the average relative error is below 1%. The inference time for a single case is 0.017 7 s, achieving a speedup of approximately 7 500 times, which is more than three orders of magnitude faster than traditional solvers. To further verify the model′s ability to restore real physical quantities, test data under five typical operating conditions are compared. The average relative error between the simulated and measured values is 13.25%, confirming the method′s advantages in handling real physical data. This study provides a surrogate modeling approach for online mapping and rapid calculation of stress fields in complex structures, achieving a balance between fidelity and computational efficiency.