Flow3DNet: A deep learning framework for efficient simulation of three-dimensional wing flow fields

Artificial intelligence is considered an effective means to accelerate the simulation of wing flow fields. However, rapid simulation of three-dimensional wing flow fields remains a challenging task. In particular, with the gradual accumulation of datasets, current deep learning models struggle to achieve adaptability to flow fields of arbitrary resolutions, both in the training and prediction phases. This limitation hinders the development of large-scale fluid dynamics models. To address this issue, this paper first establishes a fluid dataset specifically designed for the rapid simulation of three-dimensional wing flow fields. Additionally, a deep learning model architecture is proposed that is adaptable to flow fields of arbitrary resolutions. To enhance the prediction capability of the deep learning model and its adaptability to three-dimensional flow fields of varying resolutions, we further introduce a resolution and feature memory pool module. Extensive quantitative and qualitative analyses of the model's test results show that the deep learning model can generate flow field predictions four orders of magnitude faster than traditional Computational Fluid Dynamics (CFD) simulation methods. Additionally, the prediction accuracy for pressure exceeds 99.99%. This work demonstrates the application potential of artificial intelligence in accelerating the design process of aircraft.