Variable-fidelity surrogate model based on transfer learning and its application in multidisciplinary design optimization of aircraft

Abstract

Variable-fidelity surrogate models leverage low-fidelity data with low cost to assist in constructing high-precision models, thereby improving modeling efficiency. However, traditional machine learning methods require high correlation between low-precision and high-precision data. To address this issue, a variable-fidelity deep neural network surrogate model based on transfer learning (VDNN-TL) is proposed. VDNN-TL selects and retains information encapsulated in different fidelity data through transfer neural network layers, reducing the model's demand for data correlation and enhancing modeling robustness. Two case studies are used to simulate scenarios with poor data correlation, and the predictive accuracy of VDNN-TL is compared with that of traditional surrogate models (e.g., Kriging and Co-Kriging). The obtained results demonstrate that, under the same modeling cost, VDNN-TL achieves higher predictive accuracy. Furthermore, in waverider shape multidisciplinary design optimization practice, the application of VDNN-TL improves optimization efficiency by 98.9%. After optimization, the lift-to-drag ratio of the waverider increases by 7.86%, and the volume ratio increases by 26.2%. Moreover, the performance evaluation error of the model for both the initial and optimized configurations is less than 2%, further validating the accuracy and effectiveness of VDNN-TL.