Few-shot training for flowfield reconstruction in isolators enhanced by transfer learning

This study adopts the transfer learning (TL) paradigm to address the constraints posed by data scarcity on flowfield reconstruction models, which takes wall pressure and the current operating-condition parameters as inputs, and outputs the corresponding schlieren image. The model was pre-trained on data-rich operating conditions (source domain), and few-shot training was conducted with 50 samples from a different operating condition (target domain), where the shock structures differ significantly from those in the source domain. The proposed model features an architecture that integrates enhanced multi-scale convolutional residual blocks and convolutional block attention module within generative adversarial network, and employs a generator loss function augmented with reconstruction loss. The image from generator achieved mean values of 0.76 for Structural Similarity Index Measure (SSIM), 23.16 for Peak Signal-to-Noise Ratio (PSNR), and 0.96 for Correlation Coefficient (CORR) on test set from source domain. The mean values of corresponding metrics achieved 0.71, 21.84, and 0.94 on the test set from target domain, demonstrating the robust capability in handling data scarcity through the high-quality reconstruction of shock structures that are completely different from those in source domain. A comparative analysis with current mainstream models further demonstrated the superior overall performance of the proposed model across both domains. The enhancement limit achievable through pixel-level constraints was investigated by evaluating reconstruction results under varying reconstruction loss weights.