QLingNet: An efficient and flexible modeling framework for subsonic airfoils

TL;DR

QLingNet is a novel deep learning framework that efficiently predicts subsonic flow fields at various resolutions, significantly speeding up simulations while maintaining high accuracy, especially in pressure prediction.

Contribution

The paper introduces a resolution-agnostic deep learning model with a multi-scale feature extractor and memory pooling, enhancing generalization and efficiency over existing methods.

Findings

Achieves three orders of magnitude speedup over CPU solvers.

Maintains high accuracy with over 99% pressure prediction.

Effectively handles flow fields of arbitrary resolutions.

Abstract

Artificial intelligence techniques are considered an effective means to accelerate flow field simulations. However, current deep learning methods struggle to achieve generalization to flow field resolutions while ensuring computational efficiency. This paper presents a deep learning approach for rapid prediction of two types of subsonic flow fields with different resolutions. Unlike convolutional neural networks, the constructed feature extractor integrates features of different spatial scales along the channel dimension, reducing the sensitivity of the deep learning model to resolution while improving computational efficiency. Additionally, to ensure consistency between the input and output resolutions of the deep learning model, a memory pooling strategy is proposed, which ensures accurate reconstruction of flow fields at any resolution. By conducting extensive qualitative and quantitative analyses on a given test dataset, it is demonstrated that the proposed deep learning model can achieve a three-order-of-magnitude speedup compared to CPU-based solvers while adapting to flow fields of arbitrary resolutions. Moreover, the prediction accuracy for pressure exceeds 99\%, laying the foundation for the development of large-scale models in the field of aerodynamics.