Flame development prediction of supersonic combustion flow based on lightweight cascaded convolutional neural network.

This article focuses on the development and evaluation of a memory fusion cascade network (MFCN) designed to predict flame luminescence images in a hydrogen-fueled scramjet combustor under multi- and long-span conditions. Using experimental data from a ground pulse combustion wind tunnel with Mach 2.5 flow, the MFCN model was trained and tested against existing models (Kongs and ResNet16), demonstrating superior prediction accuracy, stability, and significantly reduced model size (522.38 KB compared to 55 MB for Kongs). The study highlights the MFCN’s innovative cascaded architecture incorporating a memory fusion mechanism that effectively balances prediction precision and computational efficiency, providing valuable insights for active combustion control in scramjet engines. These findings contribute to advancing non-intrusive monitoring and control methods for complex supersonic combustion processes relevant to hypersonic vehicle propulsion.