Study of transfer learning from 2D supercritical airfoils to 3D transonic swept wings

Runze Li, Yufei Zhang, Haixin ChenarXiv 2022 · DOI · arXiv

Abstract

Machine learning has been widely utilized in fluid mechanics studies and aerodynamic optimizations. However, most applications, especially flow field modeling and inverse design, involve two-dimensional flows and geometries. The dimensionality of three-dimensional problems is so high that it is too difficult and expensive to prepare sufficient samples. Therefore, transfer learning has become a promising approach to reuse well-trained two-dimensional models and greatly reduce the need for samples for three-dimensional problems. This paper proposes to reuse the baseline models trained on supercritical airfoils to predict finite-span swept supercritical wings, where the simple swept theory is embedded to improve the prediction accuracy. Two baseline models for transfer learning are investigated: one is commonly referred to as the forward problem of predicting the pressure coefficient distribution based on the geometry, and the other is the inverse problem that predicts the geometry based on the pressure coefficient distribution. Two transfer learning strategies are compared for both baseline models. The transferred models are then tested on the prediction of complete wings. The results show that transfer learning requires only approximately 500 wing samples to achieve good prediction accuracy on different wing planforms and different free stream conditions. Compared to the two baseline models, the transferred models reduce the prediction error by 60% and 80%, respectively.

Introduction: The Curse of Dimensionality in 3D Aerodynamics

In aerospace engineering, Computational Fluid Dynamics (CFD) is essential for design, yet it faces a significant "curse of dimensionality." While data for 2D airfoils is relatively abundant and cheap to generate, 3D wing simulations are exponentially more complex and computationally expensive. A 3D wing introduces complex phenomena such as swept angles, finite spans, and interference effects that are absent in 2D.

Comparison of complexity and data requirements across different aerodynamic scales.

Training a deep learning model from scratch for 3D problems typically requires thousands of high-fidelity samples. This research explores whether we can leverage existing knowledge from 2D airfoil models to efficiently predict 3D wing performance with minimal additional data.

Physics-Embedded Transfer Learning Framework

The core of this study is a "physics-embedded" transfer learning framework. The approach involves three main components:

  1. Baseline 2D Model: A conditional encoder-decoder model was first trained on over 11,000 2D supercritical airfoil samples. This model was designed to solve both the forward problem (geometry to pressure distribution) and the inverse problem (pressure distribution to geometry).
  2. The Physics Bridge: To bridge the gap between 2D and 3D, the authors utilized Simple Swept Theory (SST). SST provides a mathematical framework to transform 3D wing section properties into equivalent 2D values, effectively acting as a physics-based pre-processor.
  3. Transfer Learning Strategy: The weights of the 2D baseline model were frozen, and a small "Additional Neural Network" (NN) was appended. This new layer was trained to learn the "residual" 3D effects—the complex physical interactions that SST cannot capture, such as flow at the wing-body junction.
The proposed transfer learning framework bridging 2D airfoils and 3D swept wings.

The researchers tested various strategies for placing this additional NN, ultimately finding that correcting the pressure coefficient distribution ($C_p$) was more effective than correcting geometric parameters.

Data and Results

The models were validated using a 3D dataset of 5,000 wing samples based on the NASA Common Research Model (CRM). The results demonstrate remarkable data efficiency: the transferred model reached a performance plateau with only 500 3D samples, a fraction of what would be needed for a standalone 3D model.

Key performance metrics include:

  • Error Reduction: The transferred models reduced prediction errors by 60% for the forward problem and 81% for the inverse problem compared to using the 2D baseline alone.
  • High Accuracy: The average Root Mean Square Error (RMSE) for 3D wing pressure predictions was just 3.08%.
  • Visual Fidelity: Qualitative analysis confirmed that the model accurately captures critical features like shock wave locations and spanwise continuity across various flight conditions.

Conclusion and Future Work

This study proves that physics-embedded transfer learning is a powerful tool for high-dimensional aerodynamic modeling. By using classical theory to handle the bulk of the transformation, neural networks can focus on learning the nuanced residuals of 3D flow. Future work will aim to extend this methodology to more complex aircraft configurations, including those with engine nacelles and non-planar wing geometries, and will utilize higher-fidelity RANS solvers for 3D sample generation.

Summary updated

Abstract

Machine learning has been widely utilized in fluid mechanics studies and aerodynamic optimizations. However, most applications, especially flow field modeling and inverse design, involve two-dimensional flows and geometries. The dimensionality of three-dimensional problems is so high that it is too difficult and expensive to prepare sufficient samples. Therefore, transfer learning has become a promising approach to reuse well-trained two-dimensional models and greatly reduce the need for samples for three-dimensional problems. This paper proposes to reuse the baseline models trained on supercritical airfoils to predict finite-span swept supercritical wings, where the simple swept theory is embedded to improve the prediction accuracy. Two baseline models for transfer learning are investigated: one is commonly referred to as the forward problem of predicting the pressure coefficient distribution based on the geometry, and the other is the inverse problem that predicts the geometry based on the pressure coefficient distribution. Two transfer learning strategies are compared for both baseline models. The transferred models are then tested on the prediction of complete wings. The results show that transfer learning requires only approximately 500 wing samples to achieve good prediction accuracy on different wing planforms and different free stream conditions. Compared to the two baseline models, the transferred models reduce the prediction error by 60% and 80%, respectively.
BibTeX
@misc{li2022study,
  author = {Runze Li and Yufei Zhang and Haixin Chen},
  title = {{Study of transfer learning from 2D supercritical airfoils to 3D transonic swept wings}},
  year = {2022},
  eprint = {2206.02625},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.02625},
}
Figure