Fast prediction of three-dimensional wing inviscid flow fields from sparse data based on an autoencoder and a deep attention network
Authors:
- Mingrui LiMLMingrui Li
- Jingying WangJingying Wang
- Ke SunKSKe Sun
- Fangzhou HanFHFangzhou Han
- Chunhian LeeCLChunhian Lee
- 1 August 2025
- journal article
- Published by AIP Publishing in Physics of Fluids
- Vol. 37 (8)
Abstract
Compared to traditional computational fluid dynamics methods, which involve high computational costs to solve the governing flow equations, deep learning has emerged as a promising alternative for rapid flow field prediction. This paper introduces a novel deep learning method for quickly reconstructing three-dimensional (3D) wing inviscid flow fields under sparse data conditions. Initially, an autoencoder model is utilized to extract the geometric features of the 3D wing. This model efficiently compresses the point cloud data of the wing into a six-dimensional latent vector, facilitating a compact representation of complex geometric information. Next, a residual-attention multilayer perceptron (RAMLP) model, which incorporates attention mechanisms and residual connections, is introduced to predict the flow field variables, including velocity components in the x, y, and z directions, as well as the pressure coefficient. The model takes geometric feature parameters, Mach number, angle of attack, and flow field coordinates as inputs, allowing it to generalize effectively across various incoming flow conditions. Finally, a comparative analysis is conducted to evaluate the performance of RAMLP against two commonly used aerodynamic surrogate models: multilayer perceptron (MLP) and multi-head perceptron (MHP), in the context of 3D wing inviscid flow field prediction. Although RAMLP has a more complex computational structure, it maintains efficient prediction times, processing a single condition in under one second. Furthermore, RAMLP exhibits superior prediction accuracy under sparse data conditions compared to both MLP and MHP models.Keywords
Funding Information
- National Natural Science Foundation of China (12002193)
- Future Project for the Young Scholars of Shandong University (31380082164048)
This publication has 33 references indexed in Scilit:
- A robust high-resolution discrete-equations method for compressible multi-phase flow with accurate interface capturingJournal of Computational Physics, 2023
- A parametric study of rectangular jets issuing into a laminar crossflowExperiments in Fluids, 2023
- Physics-informed deep learning for simultaneous surrogate modeling and PDE-constrained optimization of an airfoil geometryComputer Methods in Applied Mechanics and Engineering, 2023
- Flow field modeling of airfoil based on convolutional neural networks from transform domain perspectiveAerospace Science and Technology, 2023
- A general deep transfer learning framework for predicting the flow field of airfoils with small dataComputers & Fluids, 2022
- Fast sparse flow field prediction around airfoils via multi-head perceptron based deep learning architectureAerospace Science and Technology, 2022
- An enhanced hybrid deep neural network reduced-order model for transonic buffet flow predictionAerospace Science and Technology, 2022
- A deep learning approach for the transonic flow field predictions around airfoilsComputers & Fluids, 2022
- Fast pressure distribution prediction of airfoils using deep learningAerospace Science and Technology, 2020
- Prediction of aerodynamic flow fields using convolutional neural networksComputational Mechanics, 2019