Abstract: Due to the complexity of turbulent flow problems for film cooling, the traditional Reynolds average Navier-Stokes (RANS) method tends to underestimate the intensity of turbulent thermal diffusion, leading to inaccurate prediction of cooling effectiveness. A framework based on physics-informed neural network (PINN) was therefore proposed, and a data-driven neural network model of turbulent Prandtl number was built based on RANS flow data and large eddy simulation(LES) temperature data. After implementing this model into a RANS solver, the intensity of turbulent thermal diffusion could be adjusted dynamically and a temperature distribution highly consistent with LES results was obtained. Results show that PINN is an effective method to build a data-driven turbulence model and modeling of turbulent Prandtl number can effectively improve the accuracy of RANS temperature prediction.