Abstract: The flow of liquid lead-bismuth coolant in fuel assemblies exhibits forced convection, natural convection, and mixed convection states where both coexist. This flow and heat transfer state directly influences the heat transfer efficiency of the coolant and the core temperature distribution, which is crucial for the thermal-hydraulic safety of the core and the overall operational reliability. This study focused on two-dimensional backward-facing step of liquid lead-bismuth and developed a mixed convection heat transfer physics-informed neural network (MCHT-PINN) framework based on the adaptive weighting GradNorm algorithm. The model integrates governing continuity equation, Navier-Stokes (N-S) equations and energy equation, which are incorporated directly into the loss function alongside sparse high-fidelity data from OpenFOAM simulations. By integrating physics constraints with sparse data-driven learning and dynamically balancing multi-task loss functions, the model achieves high-accuracy prediction of thermo-hydraulic parameters and elucidates the mechanisms of mixed convection heat transfer in liquid lead-bismuth. As a result, the MCHT-PINN model excels in capturing flow and temperature field characteristics, exhibiting excellent agreement with OpenFOAM reference solutions and an average relative L2 error below 3%. This not only validates the model’s high accuracy and generalization capability but also highlights the effectiveness of the GradNorm adaptive weighting algorithm in dynamically balancing gradients of partial differential equations (PDE), boundary conditions, and data losses. As the Richardson number (Ri) increases, the buoyancy term gains greater weight in the momentum equation, thereby enhancing the buoyancy effect. Near the wall, buoyancy-driven secondary flows emerge, disrupting the original boundary layer. This leads to a thinning of the velocity boundary layer, while the thermal boundary layer is lifted, impeding effective heat transport from the heated wall to the mainstream region. Consequently, heat accumulates in the lower near-wall region, resulting in localized temperature elevation. This study provides a reliable predictive tool for the thermal-hydraulic characteristics of lead-based fast reactors, but its MCHT-PINN framework can also be extended to complex geometries such as three-dimensional assembly channels and rod bundles with flow disturbances, further reducing reliance on high-cost CFD data. And this research paves new pathways for the application of data-physics fusion approaches in the investigation of flow and heat transfer in liquid metals.