Flame structure reconstruction in a supersonic combustor based on lightweight contextual encoding coordinate attention network

Accurately predicting flame structure in scramjet combustors is crucial for efficient combustion and flow control in supersonic vehicles. In this study, we conducted experiments on flame structure evolution in a supersonic combustor under a 2.5 Mach inflow condition. We collected a dataset containing flame field evolution images and combustor wall pressure data. To reconstruct the flame structure, we proposed contextual encoding coordinate attention network (CECAN), a novel approach utilizing only wall pressure data. We compared CECAN's reconstruction performance with multi-view domain adaptive generative network (MV-DAGN) and ResNet16 models using evaluation criteria over short and long time spans. CECAN demonstrated superior accuracy and stability in reconstructing dynamic flame field evolution images. In the independent test set, CECAN achieved significant improvements compared to the other models, with up to 41.8% enhancement in signal-to-noise ratio, 45.69% enhancement in structural similarity, and 14.24% enhancement in the correlation coefficient. Notably, CECAN is lightweight, with a reduction of 97.54% in volume and 99.53% in computational parameters compared to MV-DAGN. Despite its efficiency, CECAN maintains a balanced performance in computational speed and accuracy. Our findings demonstrate the effectiveness and efficiency of CECAN in reconstructing flame field evolution images. Its superior performance and lightweight nature make it a promising approach for accurate flame structure prediction in scramjet combustors.