Flow mechanism-embedded deep learning for modeling geometric uncertainty effects on aerodynamic stability

Aerodynamic stability is a critical prerequisite for designing the next generation of high-load aero-compressors that achieve both high performance robustness and superior aerodynamics. However, the aerodynamic stability of compressors is highly vulnerable to manufacturing-induced geometric uncertainties, which significantly degrade the surge margin under off-design conditions. To efficiently quantify this complex relationship, the study proposes a novel Flow Mechanism-Informed Neural Network (FMINN) framework. The FMINN incorporates the axial momentum ratio of tip leakage flow directly into the hidden layers of a deep neural network, thereby embedding flow physics into the learning process. Geometric uncertainties are modeled based on actual blade measurements from a multi-axis rotary overspeed measurement system. Through Principal Component Analysis, high-dimensional scan data are reduced to a compact set of interpretable components without sacrificing essential geometric information. The FMINN outperforms traditional surrogate models with sparse geometric training samples, delivering an order-of-magnitude lower relative error than conventional neural networks and adaptive polynomial chaos expansions. Applied to an 8-stage axial compressor working at a medium‑to‑high design rotational speed, FMINN-based uncertainty quantification reveals a pronounced non-Gaussian distribution in stability margin variation, indicating a 1.14% probability of severe aerodynamic degradation. Further sensitivity analysis using Shapley Additive exPlanations quantifies how key geometric deviation features independently and interactively influence aerodynamic stability in a nonlinear manner. Flow mechanism analysis further reveals how specific deviations propagate uncertainties into the flow field. The findings of this study aim to drive the intelligent transformation of aero-compressor design from an “experience-driven” to a “model-driven” paradigm.