The performance of film cooling is influenced by many parameters, and the nonuniform flow caused by the internal cooling system is found to largely affect the film cooling, which further complicates the in-hole flow and draws new difficulties in predicting the cooling performance. In this study, we find a very interesting phenomenon that there always exists an in-hole interface, on which distributions of many parameters, including the velocity and kinetic energy, are seldom affected by the mainstream. The existence of this specific interface can be observed for both cylindrical and shaped film cooling holes under most operating conditions. The theoretical analysis of this interface is conducted in this study based on the characteristic decomposition of the Navier–Stokes equation, and this interface is named as the characteristic interface. Theoretical analysis and numerical observations suggest the film cooling system can be simplified to two weakly coupled regions separated by this interface. It also explains why existing source term models for film cooling may fail. Based on these findings, a new prediction model is developed, which uses the convolutional neural networks (CNN) model to predict the boundary conditions on the characteristic interface. The new model outperforms existing source term models and yields similar accuracy as full-mesh computational fluid dynamics (CFD), while reducing the computational cost by one order of magnitude. This model is further evaluated in large eddy simulation (LES), showing moderate success. To sum up, the current work reports the characteristic interface phenomenon in the film cooling hole, based on which a new and efficient prediction model is developed and verified.