Abstract
Aim/Background: This study presents a hybrid Physics-Informed Neural Network (PINN) framework for solving nonlinear partial differential equations (PDEs) encountered in computational fluid dynamics (CFD), with specific application to the thermal-hydraulic analysis of nuclear reactor fuel cladding. Reliable CFD modeling is critical in nuclear engineering because fuel cladding operates under severe heat flux, high-pressure coolant flow, and irradiation-induced material degradation, all of which strongly influence reactor safety and performance.
Method: The proposed framework combines deep neural networks with physical constraints derived from the Navier-Stokes equations, while simultaneously enforcing boundary and initial conditions. Sparse observational data are incorporated through a multi-component adaptive loss function to improve predictive accuracy and stability. The model was evaluated using benchmark CFD problems, including Burgers’ equation and the lid-driven cavity flow at Reynolds numbers ranging from 50 to 3200. Results were compared with established numerical solutions obtained from the finite element method (FEM) and finite volume method (FVM).
Results: The hybrid PINN achieved a relative L² error of 0.63% for velocity field prediction at Reynolds number Re = 100, substantially outperforming conventional PINNs, which recorded 4.21% error. In addition, the framework reduced computational cost by 93.5% compared with FEM during inference. This level of efficiency supports near real-time prediction of coolant flow and heat transfer behavior around Zircaloy and SiC/SiC composite cladding geometries. Ablation studies and hyperparameter sensitivity analysis further confirmed the importance of each architectural component.
Conclusion: The findings demonstrate that hybrid PINNs are accurate, computationally efficient, and physics-consistent surrogate models for benchmark CFD problems. While the validation is conducted on standard test cases (lid-driven cavity and Burgers’ equation), the framework is designed with direct applicability to nuclear reactor fuel cladding thermal-hydraulics, where its mesh-free and data-assimilation capabilities address key challenges in subchannel and accident analysis. Full nuclear cladding validation against experimental datasets remains the subject of ongoing work.

Key words: Adaptive loss weighting, finite element methods, fuel cladding, neural networks, scientific machine learning, Zircaloy cladding.