Deep learning methods for super-resolution reconstruction of temperature fields in a supersonic combustor

Abstract

A general super-resolution (SR) reconstruction strategy is proposed to address the super-resolution reconstruction of temperature fields from low-resolution coarse temperature field data using convolutional neural networks. Two deep learning (DL) models were applied to augment the spatial resolution of temperature fields. One is the classical super-resolution convolutional neural network, and the other is the novel multiple path super-resolution convolutional neural network (MPSRC). Three paths with and without a pooling layer are designed in the MPSRC to fully capture spatial distribution features of temperature. Numerical simulations of combustion in a strut scramjet combustor at various Mach numbers are carried out to establish a dataset for network training and testing. The corresponding high-resolution temperature fields were successfully reconstructed with remarkable accuracy. The reconstruction performances of those models were comprehensively investigated and compared with the bicubic interpolation method. The results demonstrated that both DL methods can greatly improve the super-resolution reconstruction accuracy and the MPSRC can provide a better reconstruction result with a lower mean square error and a higher peak signal-to-noise ratio.