1 April 2025 in Physics of Fluids
This paper constructs a physics-informed neural network with boundary identification strategy (BI-PINN) to reconstruct the flow field of an ultra-high-load, low-pressure turbine cascade using sparse data. Since the boundary layer position is unknown before training, an iterative method is required to determine the boundary layer range. The BI-PINN trains the network based on a dual convergence criterion: achieving flow field reconstruction convergence in both the boundary layer and mainstream regions to extract the current boundary layer range and then ensuring the convergence of the boundary layer range. To improve neural network convergence while maintaining high training speed, this study assigns different weights to the governing equation terms in the loss function for the boundary layer and mainstream regions. The BI-PINN method demonstrates computational accuracy comparable to computational fluid dynamics in inverse problems with incomplete boundary conditions. This study finds that for high-load, low-pressure turbine cascades, selecting pressure observations only on the blade surface is sufficient for high-accuracy flow field reconstruction. However, for ultra-high-load, low-pressure turbine cascades, this paper proposes a method of selecting multiple normal-direction pressure observations along the blade surface to achieve high-accuracy flow field reconstruction with a small observation set. The BI-PINN method balances the conflict between the low-gradient mainstream region and the high-gradient boundary layer region in flow field reconstruction. By leveraging both physical governing equations and observation data, it enables precise full flow field reconstruction of ultra-high-load, low-pressure turbine cascades using a small dataset and a neural network with a limited number of parameters.
1 November 2025 in Physics of Fluids
The cross-flow effect in the compressor corner region alters the separation-induced transition flow of the suction surface boundary layer. As the Reynolds number (Re) decreases, the mechanism of three-dimensional boundary layer separation-induced transition flow coupled with crossflow becomes more complex, making the analysis of loss formation mechanisms more difficult, thereby heightening the challenge of efficient flow control. This study investigates a high-subsonic compressor cascade to elucidate the fundamental mechanisms by which Re effects induce variations in separation-induced transition characteristics across different blade heights, with particular emphasis on the coupling mechanism between crossflow and streamwise instability. The results demonstrate that crossflow induces spanwise non-uniformity in the separation, transition and reattachment positions of the three-dimensional suction surface boundary layer. Elevated viscosity, identified as the primary factor delaying shear layer transition at low Re, retards the growth of maximum streamwise velocity differential near the separated shear layer. This attenuation reduces the growth rate of Kelvin–Helmholtz (K–H) instability, ultimately postponing transition. Concurrently, low-Re effect enhances the sensitivity of three-dimensional boundary layer separation-induced transition characteristics to crossflow disturbances. The axial transition position under crossflow interaction exhibits a progressive advancement from end wall to midspan followed by subsequent retardation. Within regions devoid of secondary separation, a linear correlation exists between transition location and the position of maximum spanwise velocity in separation bubbles. In contrast, transition positions show strong dependence on reverse crossflow extent in secondary separation regions. Finally, the maximum separation bubble thickness within the separation-induced transition regime determines the turbulent viscous dissipation occurring in the separated shear layer, and suppressing the crossflow should be considered to reduce the profile loss, which provides a new perspective for the efficient loss control for compressors at low Re.
1 September 2025 in Physics of Fluids
Strong three-dimensional effects within low Reynolds number compressor cascades invalidate the assumption of turbulence isotropy, significantly reducing the accuracy of solving the Navier–Stokes equations using Reynolds-averaged Navier–Stokes (RANS) methods. Although large eddy simulation (LES) maintains high accuracy at low Reynolds numbers, it comes with substantially higher computational costs. Therefore, this study integrates RANS and LES samples to construct a variable-Re multi-fidelity dataset. Based on this, a film-Re physics-guided multi-fidelity network (FP–MFN) is developed to predict flow fields at Reynolds numbers beyond those used in the training set. The FP–MFN model consists of a low-fidelity network and a high-fidelity network, which are coupled in the loss function via a fidelity fusion coefficient. When the fidelity fusion coefficient is set to 0.5, the prediction demands of the two models reach an optimal balance, yielding the highest prediction accuracy. The FP–MFN model incorporates a Reynolds number modulation network in part of its hidden layers, enhancing its ability to learn the relationships between flow field structures and Reynolds number variations. This design substantially enhances the prediction accuracy of flow fields at Reynolds numbers unseen during training compared with conventional multi-fidelity networks. Furthermore, the FP–MFN incorporates a physics-guided loss term into the loss function, ensuring the physical consistency of predicted flow fields under previously unseen Reynolds conditions. The FP–MFN model accurately predicts flow separation in low-Reynolds-number compressor cascades under different loading conditions, highlighting its strong generalization capability.
29 August 2025 in Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering
In numerical simulations, finer grids generally yield higher accuracy compared to coarse grids, but significantly more computational resources and time are demanded. In scientific researches, it is often necessary to compute many cases with similar boundary conditions. Traditional Computational Fluid Dynamics simulations solve each case individually, thereby ignoring the shared characteristics among similar cases, which leads to inefficient use of computational resources. To address the issue, this paper proposes a flow field reconstruction model for compressors based on deep neural networks combined with a fine-tuned strategy (FTDNN). This model captures common flow patterns among similar cases to enable fast and high-precision predictions for unseen cases. On a variable aerodynamic dataset, a deep neural network with two prediction modes is constructed: one is a pre-trained mode, and the other uses a fine-tuned strategy for high-accuracy predictions. In deep learning models, generalization under multiple operating conditions and precision under specific conditions often present a trade-off. This paper quantitatively analyzes the relationship between the number of frozen hidden layers and prediction accuracy. It is ultimately found that when the number of frozen and trainable hidden layers is equal, a balance between generalization and accuracy is achieved, resulting in optimal prediction performance. The FTDNN model achieves a maximum average relative error of only 1.29% on extreme flow fields near the upper limit of positive attack angles in the test set. The model also demonstrates excellent accuracy in predicting flow fields at the compressor blade leading edge, within boundary layers, and in wake distributions.
7 March 2026 in Aerospace Science and Technology
25 August 2026 in Aerospace Science and Technology