Rapid evaluation of turbine-vane flow fields is valuable for repeated aerodynamic analyses. Efficient prediction of three-dimensional turbine-vane flow fields remains challenging because of their spatial complexity and the computational cost of conventional simulations. This study combines a radial-sectional flow representation with a multi-branch, multi-scale Swin Transformer to rapidly predict six steady-RANS flow variables from operating conditions, sectional geometry, and normalized radial position. The predicted sectional fields are assembled into a structured three-dimensional representation. A CFD database for the GE E3 high-pressure turbine vane comprises 900 operating conditions and 115,200 radial sectional fields. Of these, 400 conditions are assigned to model development and 500 are reserved for external evaluation. The network is trained using 8192 sectional samples from 256 conditions, representing approximately 7.1 percent of the complete sectional database. On the 500-condition external dataset, the model achieves an average normalized error of approximately 0.22 percent and a gradient error of approximately 3x10^(-4). Comparisons of representative flow fields show errors below 1 percent over most displayed regions. The predicted fields also reproduce the principal locations and overall development trends of the dominant vortical structures. Prediction and assembly of all 128 sections, yielding approximately 4.2 million structured sampling locations, take about 9 s on an RTX 4090 GPU, compared with approximately 30 min for a representative CFD calculation on a 20-core Xeon CPU. These results support rapid aerodynamic evaluation of the investigated vane across the sampled operating range.

Rapid Prediction of Three-Dimensional Turbine Vane Flow Fields via Radial-Sectional Decomposition and a Multi-Scale Swin Transformer / Li, Z., Wen, F., Misul, D.A., Zhao, Z., Wang, S., Salvadori, S.. - In: AEROSPACE SCIENCE AND TECHNOLOGY. - ISSN 1270-9638. - ELETTRONICO. - 180:7(2026), pp. 1-17. [10.1016/j.ast.2026.114115]

Rapid Prediction of Three-Dimensional Turbine Vane Flow Fields via Radial-Sectional Decomposition and a Multi-Scale Swin Transformer

Daniela Anna Misul;Simone Salvadori
2026

Abstract

Rapid evaluation of turbine-vane flow fields is valuable for repeated aerodynamic analyses. Efficient prediction of three-dimensional turbine-vane flow fields remains challenging because of their spatial complexity and the computational cost of conventional simulations. This study combines a radial-sectional flow representation with a multi-branch, multi-scale Swin Transformer to rapidly predict six steady-RANS flow variables from operating conditions, sectional geometry, and normalized radial position. The predicted sectional fields are assembled into a structured three-dimensional representation. A CFD database for the GE E3 high-pressure turbine vane comprises 900 operating conditions and 115,200 radial sectional fields. Of these, 400 conditions are assigned to model development and 500 are reserved for external evaluation. The network is trained using 8192 sectional samples from 256 conditions, representing approximately 7.1 percent of the complete sectional database. On the 500-condition external dataset, the model achieves an average normalized error of approximately 0.22 percent and a gradient error of approximately 3x10^(-4). Comparisons of representative flow fields show errors below 1 percent over most displayed regions. The predicted fields also reproduce the principal locations and overall development trends of the dominant vortical structures. Prediction and assembly of all 128 sections, yielding approximately 4.2 million structured sampling locations, take about 9 s on an RTX 4090 GPU, compared with approximately 30 min for a representative CFD calculation on a 20-core Xeon CPU. These results support rapid aerodynamic evaluation of the investigated vane across the sampled operating range.
File in questo prodotto:
File Dimensione Formato  
1-s2.0-S1270963826024910-main_compressed-1.pdf

accesso riservato

Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Non Pubblico - Accesso privato/ristretto
Dimensione 986.94 kB
Formato Adobe PDF
986.94 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/3017393