This study proposes a multifidelity radial basis function neural network (MFRBFNN) model and an associated surrogate optimization framework to facilitate the efficient aerodynamic shape optimization of aircraft. Analytical benchmark functions demonstrate that the MFRBFNN model has the ability to capture linear correlations across multifidelity samples at a level that is comparable to the Co-Kriging model and superior to the multifidelity deep neural network (MFDNN) model. Furthermore, the MFRBFNN model effectively captures nonlinear correlations that the Co-Kriging model fails to extract while outperforming the MFDNN model. As a result of its gating network, the MFRBFNN model offers higher tuning efficiency and greater model robustness compared to the MFDNN model. Finally, model-agnostic infill sampling criteria and a MFRBFNN-based surrogate optimization framework are established. In unconstrained RAE2822 airfoil optimization and multiconstrained high-dimensional wing optimization cases, the proposed method achieves 16.61% and 1.68% performance improvements over traditional methods, respectively, while reducing computational time by 11.39% and 29.42% compared to single-fidelity approaches. These results confirm that the MFRBFNN model and optimization framework can significantly enhance the efficiency of aerodynamic shape optimization.
Lian et al. (Tue,) studied this question.