High-fidelity prediction of LBE mixed convection heat transfer via gradient-balanced physics-informed neural networks

Mixed convection in liquid metals, such as lead-bismuth eutectic (LBE), presents intricate flow and heat transfer characteristics due to the strong coupling of forced and natural convection in low Prandtl number fluids. Accurately predicting these phenomena is paramount for the thermal-hydraulic safety and design of advanced nuclear reactors. In this study, a novel Mixed Convection Heat Transfer Physics-Informed Neural Network (MCHT-PINN) is proposed to conduct progressive case studies across a two-dimensional backward-facing step, a two-dimensional circular pipe, and a three-dimensional 19-pin fuel assembly. To address the highly coupled multi-physics fields, the MCHT-PINN architecture integrates the GradNorm adaptive-weight algorithm, which dynamically balances the loss gradients associated with the Partial Differential Equations (PDEs) and boundary conditions. The results demonstrate that the MCHT-PINN model achieves accurate predictions of both velocity and temperature fields, demonstrating excellent agreement with OpenFOAM reference solutions. The average relative L 2 errors are consistently maintained below 3% across all cases. Furthermore, the model accurately resolves complex coupled thermo-fluid phenomena, capturing not only buoyancy-driven velocity profile distortions but also the localized lifting of the thermal boundary layer and subsequent flow field evolution. Ultimately, these findings not only verify the robustness of the MCHT-PINN across varying geometries but also provide a reliable computational tool for evaluating complex thermal-hydraulics in liquid metal reactor components.