A Convolutional Neural Network to Predict the Mass Flow Rates of the Holes in Double-Wall Flat Plate

In the cooling design of double-wall turbine blades, it is crucial to accurately evaluate the coolant mass flow rates of each impingement hole and film hole as it directly affects the integrated heat transfer effect. CFD method requires significant computational resources, while flow network method struggle to accurately model interconnected channels and complex migration flows. Therefore, the convolutional neural network is introduced in present study to predict the mass flow rates of the double-wall flat plates with different number of holes and pressure gradients on the film hole outlets. The pressure differences of each hole are spliced together to form a one-dimensional tensor as the input, with the mass flow rates of each film hole and impingement hole as the outputs. The mass conservation is added to the loss function as a regularization term to constrain network training. The model is trained and tested by simulation data, and it achieves the mean relative errors of 1.96% and 3.01% for the film and impingement holes in the test dataset, respectively. An optimization is introduced by adding an edge identification tensor into the input tensor and a corresponding convolutional filter input channel to address the problem of large prediction errors for holes at edge. Compared to the baseline model, the optimized model reduced the mean relative errors for all impingement holes and film holes at edge in the test dataset by 40.69% and 40.75%, respectively.