Under typical operating conditions within a pressurized water reactor (PWR) core, the Reynolds number reaches the order of 10⁵, characterizing the coolant flow with pronounced, highly nonlinear dynamics. When computationally resolving such flow fields, an inherent physical mismatch arises between the actual, complex flow states and boundary conditions, and the idealized conventional governing equations. Within the context of deep learning frameworks, this discrepancy engenders severe conflicts between the data-fitting objectives and the physical-equation constraints, ultimately leading to convergence failures and suboptimal solution accuracy. To address this fundamental challenge, this paper proposes a novel computational approach: the Spatially Adaptive Physics-Informed Neural Network (SA-PINN). The SA-PINN model introduces a spatially varying adaptive tuning factor into the governing equations. By leveraging the network's dynamic learning capabilities, it iteratively refines the idealized physical model to compensate for and assimilate unmodeled, real-world flow effects. Empirical results demonstrate that this mechanism effectively resolves the convergence impediments stemming from data-equation inconsistencies. Relying on a mere 20% sparse data sampling, the SA-PINN achieves high-fidelity flow field reconstruction. Notably, in regions characterized by intense turbulent mixing, the mean squared error (MSE) of the velocity field is reduced to 1.94×10^-3, with the coefficient of determination (R²) exceeding 0.958. This research advances an efficient, highly adaptable methodological framework for reconstructing complex flow fields in PWR rod bundles under conditions of extreme data sparsity.
 

Physics-informed neural networks; Spatially adaptive correction; Rod bundle; Sparse data;