Artificial intelligence (AI) and machine learning (ML) have profoundly revolutionized the aviation industry, and researchers are increasingly prioritizing these new technologies in various applications. However, the traditional end-to-end (E2E) neural networks suffer from insufficient feature extraction from single-source input data, which limits the model’s performance. To overcome this drawback, a framework of an adversarial feature fusion network (AFF-Net) is proposed and applied to the design space mapping of fans. In this framework, the flowfields are regarded as heterogeneous forms of fan performance metrics and are utilized to fully exploit the latent features. Moreover, a feature augmentation technique and adversarial training strategies are employed to alleviate the irregularities and discontinuities of the latent feature manifold and enhance the effects of feature fusion. Compared to the traditional E2E model, the prediction errors of fan efficiency and total pressure ratio are reduced by 22 and 24%, respectively. Finally, two engineering applications are demonstrated, which aim to show the engineering application potential of AFF-Net and investigate the influence of different design parameters and their combinations on fan performance. This study aims to fill the gap in the design space mapping of low-speed axial fans and provide guidance for fan blade design.
Ma et al. (Thu,) studied this question.