Multi-fidelity neural network for instantaneous aerodynamic prediction of three-dimensional flapping wings

Constructing surrogate models that can quickly predict the instantaneous aerodynamic performance of flapping wing motions is crucial for the optimization of flapping wings. Machine-learning-based surrogate models require extensive computational resources to generate sufficient training data, which limits their practical application. In this work, a multi-fidelity neural network framework is proposed to predict the instantaneous three-dimensional (3D) aerodynamic performance in flapping wing motions. First, based on the samples of simplified two-dimensional flapping wing motion, a convolutional neural network model is constructed to generate pseudo-3D data as low-fidelity (LF) data. Then a model based on a gated recurrent unit and a Transformer encoder is constructed, which combines the LF data and the high-fidelity data obtained from 3D simulation. Compared to the conventional method, the multi-fidelity neural network model reduces the prediction error by over 30% and decreases the data generation time by 27.22%. This work provides a computationally efficient framework for aerodynamic performance prediction in 3D flapping wing, which is potential in the design and optimization of flapping wing vehicles.