A study on flame reconstruction in a supersonic combustor using deep learning

This study investigates the application of a low-order reconstruction method for image reconstruction of a scramjet combustor. In the encoding network, reconstruction performance was assessed by evaluating adjustments to sampling channel count and modifications to neural network architectures. Upsampling methods such as convolutional neural networks (CNNs), interlayer attention mechanisms, and pixel shuffle were tested in the decoder network. Furthermore, a parameter expansion strategy based on the enlargement of convolutional feature map channels was proposed and examined. The results were quantified by morphological and frequency domain analyses under tests with datasets of different equivalence ratios, suggesting the effectiveness of the scheme for flashback prediction. It was found that the reconstruction effect of 6-point sampling is close to that of continuous sampling (68 points), which is the most cost-effective among the tested schemes. By comparing different network structures, the method proposed in this paper achieves better reconstruction results than the large-parameter CNN network with a small-scale network structure.