Variable fidelity aerodynamic modeling method based on transfer learning

The establishment of the aerodynamic model based on discrete data sets is an important part of aircraft optimization design. However, it takes a long time and high cost to build a complete and high-precision numerical simulation and wind tunnel test data set. In order to shorten the development cycle and save the design cost, this paper proposes a variable fidelity aerodynamic modeling method based on transfer learning, aiming at establishing high-precision aerodynamic models on limited data sets. This method combines the aerodynamic data fusion theory and transfer learning method, designs a regression network structure based on Long Short-Term Memory (LSTM) neural network, and adopts the parameter tuning mechanism of pre-training and fine-tuning for transfer training, so as to obtain the aerodynamic model with high fidelity. Specifically, taking the XFLR calculation data (low precision) and wind tunnel test data (high precision) of NACA 2414 airfoil as the research object, a high-fidelity aerodynamic prediction model is formed by using a small amount of high-precision data to conduct transfer learning on the pre-trained model of a large number of low-precision data sets. Then, a wind tunnel data modeling experiment with data volume ranging from 1/2 to 1/10 was designed, and it was compared with the unmigrated LSTM model and the Additive Scaling Function Based Multi-fidelity Surrogate (AS-MFS) model. Experimental results show that the proposed method achieves higher prediction accuracy under all data quantities, and the prediction accuracy of the drag and lift-drag ratio is increased by 7.22% and 8.85% on average, respectively, compared with that before migration; compared with AS-MFS, the average improvement is 8.66% and 4.36%, respectively.