Session: 36-04 UQ & Sensitivity Analysis - Part 1

Submission Number: 176177

Uncertainty Quantification of Geometric Deviations in Subsonic Cascades at Low Reynolds Numbers Using Multi-Fidelity Neural Networks 

Geometric deviation uncertainty significantly impairs the aerodynamic performance of aero-engines. In the context of refined design, the aerodynamic robustness of compressor blades is highly susceptible to manufacturing deviations. However, due to the high-dimensional nature of their design parameters, traditional uncertainty quantification (UQ) methods, such as the non-intrusive polynomial chaos expansion method (NIPCE), suffer from the curse of dimensionality. This is because the computational cost of constructing orthogonal polynomial expansions increases significantly with the growth of the dimension of the stochastic space. With the advancement of deep learning techniques, high-dimensional nonlinear surrogate models have provided new avenues for alleviating the curse of dimensionality. Among them, the multi-fidelity neural network (MFNN) has emerged as a research focus in recent years. By integrating low-cost, low-fidelity samples with high-cost, high-accuracy samples, MFNN can substantially reduce the reliance on expensive high-fidelity evaluations while maintaining predictive accuracy. However, existing MFNNs still exhibit notable limitations. Their learning paradigm is essentially a “point-to-point mapping”, enabling feature extraction from low-fidelity data only at locations corresponding to available high-fidelity samples. As a result, the large number of unmatched low-fidelity samples is merely used to construct the low-fidelity model itself, without providing effective information for correcting errors or enhancing the predictive capability of the high-fidelity model. This leads to an inefficient utilization of computational resources.

 

To address the aforementioned limitations in MFNN, this study proposes a cross-attention-enhanced multi-fidelity neural network (CAE-MFNN), which aims to improve sample learning efficiency through effective cross-fidelity feature interaction. Specifically, it calculates the attention weights between each high-fidelity sample and all low-fidelity samples, thereby enabling a “point-to-global” feature learning process. Compared with MFNN, CAE-MFNN achieves a 43% reduction in root-mean-square error and a 0.03 improvement in the coefficient of determination on benchmark test cases, while maintaining high efficiency and robustness in high-dimensional applications.

 

Although extensive research has been conducted on uncertainty quantification of geometric variations in subsonic cascades, detailed evaluations under low Reynolds number (Re) conditions remain limited. In this study, a subsonic cascade with a representative separated-bubble flow feature is selected as the research subject. A geometric deviation model is constructed based on experimental measurements, and the influence of geometric deviations on flow characteristics under different Reynolds numbers (Re = 1×10⁵–1×10⁶) is investigated using CAE-MFNN. The results show that low Re conditions exacerbate the nonlinear impact of geometric deviations on cascade aerodynamic performance. Compared with the baseline, the average total pressure loss increases by 3%, and the underlying mechanism is primarily the variation in the position and length of separation bubbles caused by geometric deviations.. These findings provide a quantitative basis for the robust design of subsonic cascades under low Re operating conditions and offer direct engineering relevance for geometric tolerance allocation and performance fluctuation control in compressors designed for high-altitude, long-endurance unmanned aerial vehicles.

Presenting Author: Wuan Zhao beihang university

Presenting Author Biography: a PhD candidate at the School of Energy and Power Engineering, Beihang University (BUAA), with research interests centered on compressor uncertainty quantification and robustness optimization, as well as low Reynolds number effects.

Authors:

Wuan Zhao beihang university
Jiang Chen Beihang University
Yi Liu Nanjing Engineering Institute of Aircraft Systems (AVIC)
Bin Li Beihang University
Hang Xiang Beihang University
Xiaodong Ren Department of Energy and Power Engineering, Tsinghua University
Chunwei Gu Department of Energy and Power Engineering, Tsinghua University