Spatial Aerodynamic Parameters Prediction for Wide-Incidence Turbines Using Small-Scale Experimental Data-Driven GNNs

Abstract Accurate performance prediction across a wide operating range remains challenging for high-performance power turbine blades. To address this issue, a graph attention network model, GAT-5s, was developed using small-scale experimental data to predict aerodynamic parameters of rotor blade profiles over a wide incidence range. On the validation profile, GAT-5s achieved mean relative errors of 1.6% for the total pressure loss coefficient, 0.2% for the outlet Mach number, and 2.7% for the deviation angle. These errors were lower than those of the back-propagation neural network baseline and the GAT-5s-CFD model trained on RANS under the same tested conditions. However, direct use of GAT-5s for surface pressure prediction produced training and testing losses above 1×10−3, indicating limited capability to describe spatial pressure distributions. To address this limitation, GAT-5G was developed by integrating pressure-point position information and a BIGRU module into GAT-5s. The model reduced the training and testing MSE losses to below 1.0×10−4, and the prediction errors for both the test and validation profiles remained within 10%. Although GAT-5G did not always yield the minimum local error among all models, it provided balanced agreement with experimental pressure distributions across the tested incidence range. For wake-loss prediction on the validation profile, the MRE also remained below 10% for the tested incidence cases. These results indicate that GAT can serve as useful surrogate models for aerodynamic prediction of wide-incidence rotor blade profiles.