Corrected Multi-Fidelity Surrogate Model Based on Deep Neural Networks for Predicting Aerodynamic Loads of Missile

Abstract Before conducting wind tunnel tests of a missile, it is important to estimate the approximate aerodynamic loads on the test model and select test equipment with appropriate size and specifications. The use of inappropriate test equipment may lead to inaccurate results and, in extreme cases, cause damage to the equipment and facility. Using low-fidelity (LF) models, which provide quick results but with lower accuracy, for missile aerodynamic load analysis may lead to an underestimation of the actual loads. On the other hand, high-fidelity (HF) models suffer from long computation times. In this study, we developed a multi-fidelity (MF) model that combines the advantages of both HF and LF models for missile aerodynamic load prediction using deep neural networks (DNNs). Among various aerodynamic forces and moments, this study focuses on normal force, axial force, and pitching moment, as these three components play a critical role in evaluating missile stability, control characteristics, and structural integrity under aerodynamic loading. The MF model requires the number of input variables for LF and HF datasets to be the same, but the conventional MF model has a limitation in this regard. We propose a novel corrected multi-fidelity (CMF) model that addresses this issue by modifying the MF model using outputs from the LF model to account for the excluded input variables from the HF dataset. The CMF model was successfully constructed using only LF data and the limited input variables from the HF dataset. As a result, the MF model showed improved accuracy compared to both HF and LF models, and the CMF model demonstrated similar accuracy to the MF model while significantly reducing the computational cost for HF sample point simulations.