Combustion flow field reconstruction in a hydrogen-fueled scramjet combustor based on lightweight generative adversarial networks

The flow characteristics of scramjets are crucial for their overall performance, necessitating the precise and efficient reconstruction of the combustor's flow field for accurate predictions. This paper tackles the challenges posed by the multitude of parameters and lengthy inference times associated with existing deep learning models by introducing a lightweight architecture specifically designed for supersonic combustor flow reconstruction. By combining the image generation capabilities of generative adversarial networks with the sophisticated feature extraction provided by a multi-head attention mechanism, the proposed model significantly enhances inference speed while preserving high performance. Validation was performed using a high-fidelity dataset derived from hydrogen combustion experiments conducted in a direct-connect pulse combustion wind tunnel, at a constant Mach number with systematically varied equivalence ratios. The results demonstrate substantial improvements, achieving a Structural Similarity Index Measure (SSIM) of 0.554 and a Peak Signal-to-Noise Ratio (PSNR) of 19.972, all while reducing the parameter count by 95 % and boosting inference speed by 56.58 % compared to the previously developed Multi-Branch Fusion Convolutional Neural Network (MBFCNN) model.