Multi-fidelity aerodynamic reduced-order model based on gating mechanism

Due to the efficiency limitations of aerodynamic data acquisition methods in engineering design, the distribution of multi-fidelity data samples is inconsistent. This results in varying accuracy levels of the constructed aerodynamic model across different regions of the design space, posing challenges for achieving global multi-fidelity data fusion. To fully utilize the imbalanced and insufficient multi-source aerodynamic data and achieve the construction of a global high-angle-of-attack unsteady aerodynamic model, a multi-fidelity model fusion and switching method based on a neural network gating mechanism is proposed. First, a Gated Switching Neural Network (GSNN) framework is constructed, using a gate unit to control the proportion of high- and low-fidelity model information. Then, a loss function regularization design is introduced, treating the low-fidelity model as the base model and achieving domain incremental learning of GSNN using a small number of high-fidelity sample points. Finally, the output of the gate unit is quantified and interpreted from the perspective of variation patterns in the physical parameters. Research results on aerodynamic modeling and flight simulation for typical aircraft indicate that the GSNN model can effectively degrade to the low-fidelity model, thereby enhancing model reliability in fully extrapolated regions. The results of unsteady aerodynamic modeling indicate that the proposed method achieves an average prediction error of approximately 6% in the test cases and provides a more accurate prediction of static aerodynamic variation. In conclusion, this paper provides a new engineering-usable technical approach for constructing global aerodynamic models of aircraft, particularly suitable for integrating and switching between aerodynamic databases and neural network models.