Efficient aerodynamic coefficients prediction with a long sequence neural network

TL;DR

This paper introduces AirfoilNet, an end-to-end neural network that efficiently predicts aerodynamic coefficients from airfoil images, combining interpretability with high accuracy and fast inference.

Contribution

The paper presents AirfoilNet, a novel neural network architecture that integrates mathematical computations with deep learning for aerodynamic prediction, enhancing interpretability and efficiency.

Findings

AirfoilNet achieves high prediction accuracy.

It demonstrates strong generalization across different airfoil data.

The model provides rapid inference suitable for practical applications.

Abstract

Traditionally, deriving aerodynamic parameters for an airfoil via Computational Fluid Dynamics requires significant time and effort. However, recent approaches employ neural networks to replace this process, it still grapples with challenges like lack of end-to-end training and interpretability. A novel and more efficient neural network is proposed in this paper, called AirfoilNet. AirfoilNet seamlessly merges mathematical computations with neural networks, thereby augmenting interpretability. It encodes grey-scale airfoil images into a lower-dimensional space for computation with Reynolds number, angle of attack, and geometric coordinates of airfoils. The calculated features are then fed into prediction heads for aerodynamic coefficient predictions, and the entire process is end-to-end. Furthermore, two different prediction heads, Gated Recurrent Unit Net(GRUNet) and Residual Multi-Layer Perceptron(ResMLP), designed to support our iteratively refined prediction scheme. Comprehensive analysis of experimental results underscores AirfoilNet's robust prediction accuracy, generalization capability, and swift inference.