Abstract
Machine learning-based models provide a promising way to rapidly acquire transonic swept wing flow fields but suffer from large computational costs in establishing training datasets. Here, we propose a physics-embedded transfer learning framework to efficiently train the model by leveraging the idea that a three-dimensional flow field around wings can be analyzed with two-dimensional flow fields around cross-sectional airfoils. An airfoil aerodynamics prediction model is pretrained with airfoil samples. Then, an airfoil-to-wing transfer model is fine-tuned with a few wing samples to predict three-dimensional flow fields based on two-dimensional results on each spanwise cross section. Sweep theory is embedded when determining the corresponding airfoil geometry and operating conditions, and to obtain the sectional airfoil lift coefficient, which is one of the operating conditions, the low-fidelity vortex lattice method and data-driven methods are proposed and evaluated. Compared to a nontransfer model, introducing the pretrained model reduces the error by 30%, while introducing sweep theory further reduces the error by 9%. When reducing the dataset size, less than half of the wing training samples are need to reach the same error level as the nontransfer framework, which makes establishing the model much easier.
Introduction
Designing modern aircraft wings requires a deep understanding of complex transonic flow, where air moves both slower and faster than the speed of sound. Traditionally, engineers rely on Reynolds-averaged Navier–Stokes (RANS) simulations for high-fidelity results. However, these 3D simulations are computationally expensive, often requiring hours or even days for a single design iteration.
While machine learning (ML) offers a path toward near-instantaneous aerodynamic prediction, these models are typically "data-hungry." Generating a 3D dataset large enough to train a reliable model can be as costly as the design process itself. This paper introduces a physics-embedded transfer learning framework that overcomes this data bottleneck by "teaching" the model to leverage much cheaper 2D airfoil data to predict 3D wing performance.
Methodology
The core innovation lies in a multi-stage pipeline that combines deep learning with classical aerodynamic principles:
- Pretraining (The 2D Expert): A 1D U-Net model is trained on a massive dataset of 22,104 supercritical airfoil simulations. This model learns the "local effect"—how specific 2D shapes influence pressure and friction.
- Physics Embedding (Sweep Theory): Instead of processing raw 3D data, the framework uses classical Sweep Theory. This principle allows the decomposition of 3D wing sections into "effective" 2D airfoils, giving the model a physics-based head start.
- Fine-tuning (The 3D Correction): A 2D ResNet-based U-Net is fine-tuned on a small set of 1,842 3D wing samples. Its role is to learn the residual "3D effects," such as tip vortices and cross-flow, that the 2D model and sweep theory cannot account for alone.
To ensure practical utility, the researchers also developed a hybrid method for estimating the Spanwise Lift Distribution (SLD). By combining the low-fidelity Vortex Lattice Method (VLM) with a small auxiliary neural network for error correction, the framework maintains high accuracy across the entire wing span.
Key Results and Discussion
The results demonstrate that physics-informed transfer learning acts as a significant "data multiplier":
- Enhanced Accuracy: Compared to a standard non-transfer model, the pretrained 2D model reduced prediction error by 30%. The addition of sweep theory further reduced error by 9%.
- Data Efficiency: The framework achieved the same error levels as baseline models while requiring less than half the 3D training samples. This drastically reduces the computational overhead of model development.
- Computational Speed: Once trained, the model provides aerodynamic predictions in approximately 2 seconds. In contrast, a full 3D RANS simulation takes roughly 15 hours on comparable hardware.
- Qualitative Performance: The model successfully captured complex flow features, including shock wave positions and flow separation zones, even in challenging test cases.
Conclusion
By embedding domain-specific physics into a transfer learning architecture, this research provides a robust method for rapid aerodynamic prediction. The ability to achieve high-fidelity 3D results using limited 3D data and extensive 2D priors marks a significant step forward in aircraft design optimization. Future work will focus on extending this approach to more complex geometries, such as wings with engine nacelles and fuselage interactions.
Summary updated