Supervised learning mixing characteristics of film cooling in a rocket combustor using convolutional neural networks

Machine learning approach has been applied previously to physical problem such as complex fluid flows. This paper presents a method of using convolutional neural networks to directly predict the mixing characteristics between coolant film and combusted gas in a rocket combustion chamber. Based on a reference experiment, numerical solutions are obtained from Reynolds-Averaged Navier-Stokes simulation campaign and then interpolated into the rectangular target grids. A U-net architecture is modified to encode and decode features of the mixing flow field. The influence of training data size and learning time with both normal and re-convolutional loss function is illustrated. By conducting numerical experiments about test cases, the modified architecture and related learning settings are demonstrated with global errors less than 0.55%.