Abstract: Computational fluid dynamics (CFD) is an important tool to evaluate the performance of turbine blades and etc. in the design stage. However, the numerical simulation of turbine blades that based on CFD method can be very time-consuming, which makes it rather difficult to meet the need of rapid iteration in the design process of turbine blades. In order to evaluate the performance of turbine blades rapidly and overcome the problem of insufficient generalization ability of pure data-driven prediction models as well, inspired by the concept of physics augmented machine learning, a novel method for turbine blade flow field prediction with strong generalization ability is proposed, by combining the similarity principle with deep learning model. Taking the prediction of the isentropic Mach number distribution at the surface of turbine blades as an example, we propose to make use of the similarity principle to normalize the geometric variables and aerodynamic parameters of turbine blades, and then prepare the training sample set and train the deep learning-based prediction model in the normalized parameter space. And accordingly, a unified prediction model based deep learning can be obtained, which can quickly predict the aerodynamic performance of turbine blades that in very different geometric size and have different boundary condition values. After finishing the model training, the trained prediction model is used to predict the flow fields of the turbine blades that works under different operation condition and of different shape in normalized design space, the flow fields of real-world blades of different size/different working conditions, and the flow fields of different section profiles of GE-E³ low-pressure turbines. The results showed that the predicted results were in good agreement with the CFD evaluation results, and the averaged relative error was less than 1.0%, which verify the accuracy and generalization ability of the proposed flow field prediction model coupling the similarity principle.