Session: 34-01 AI for flow field prediction and post-processing

Submission Number: 177971

Application of Deep Attention-Driven Multi-Subdomain Learning in Compressor Cascade Flow Field Modeling 

 Subdomain decomposition methods are commonly employed to enhance the applicability and scalability of convolutional neural networks in modeling compressor cascade flow fields, particularly when dealing with O4H structured grids. However, this method suffers from an ineffective transfer of physical information between subdomains. The lack of proper communication mechanisms between adjacent subdomains can compromise the continuity of flow variables, thereby reducing the overall modeling accuracy. As a consequence, discontinuities and accumulated errors tend to emerge at subdomain boundaries, which severely degrade the global consistency and physical fidelity of the reconstructed flow field.  Although traditional subdomain filling methods are considered to mitigate this issue to some extent, they still exhibit limitations when confronted with sparse data and intense flow gradient changes. To address this limitation, this paper proposes the Deep Cross-domain Attention U-Net (DCAU-Net). Specifically, on the basis of traditional subdomain filling methods, a cross-domain attention mechanism was designed and coupled with the U-Net neural network, while a loss constraint for filling overlapping regions was incorporated into the loss function. Research findings indicate that compared to traditional subdomain filling methods, DCAU-Net achieves significant improvements in overall modeling accuracy. In particular, at the boundaries between subdomains, DCAU-Net demonstrates superior modeling performance and flow field continuity that more closely aligns with physical solution. This approach has provided a novel solution for modeling complex nonlinear turbulence in compressors.

Presenting Author: Honglin He Shanghai Jiao Tong University

Presenting Author Biography: The Presenting author is a doctoral student at the School of Aeronautics and Astronautics, Shanghai Jiao Tong University, primarily engaged in scientific research related to aerodynamics of aircraft engine compressors and the application of artificial intelligence.

Authors:

Honglin He Shanghai Jiao Tong University
Hefang Deng Shanghai Jiao Tong University
Xiang Zuo Shanghai Jiao Tong University
Songan Zhang Shanghai Jiao Tong University
Jinfang Teng Shanghai Jiao Tong University
Mingmin Zhu Shanghai Jiao Tong University