PREDICTION OF CROSSFLOW TRANSITION BASED ON DEEP NEURAL NETWORKS

This study based on deep neural networks (DNN), produces a mapping from the physical quantities such as dimensionless velocity gradient and streamwise vorticity of laminar flow field to the intermittency of cross flow transition, and obtains a new data driven transition model. The data driven transition model is coupled with the SST k-ω turbulence model, and the process of solving the transport equations is effectively simplified, which realize efficient and accurate numerical simulation of subsonic 3-D cross flow transition. The computational data of NLF(2)-0415 swept airfoil at different Reynolds numbers is used to train DNN, and two cases are used to test. The prediction accuracy of data driven transition model is similar to that of γ-Re_θ transition model. Using the data driven transition model to compute other typical examples of cross flow transition, to verify its generalization ability. For the transition locations of NLF(2)-0415 swept airfoil with different swept angles, the simulation results of data driven transition model have similar accuracy to that of γ-Re_θ transition model. Moreover, the phenomenon of transition position moving forward and then backward in the process of sweep Angle increasing from 45° to 65° can be predicted by data driven transition model. For the standard ellipsoid, although using low resolution mesh, the data driven transition model has the ability to compute the transition location, and the computed results of C_f are same to those of other transition models and experiments. The results show that coupling data driven transition model (which is obtained from the physical quantities related to cross flow transition) with SST k-ω transition model can realize the general prediction of cross flow transition. On the premise of ensuring computational accuracy, the data driven transition model requires lower resolution mesh and has higher computing efficiency.