Graph neural networks on predicting aerodynamic flow fields around airfoils

Date

2024-9

Author

Çeltik, Süleyman Onat

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

109
views

427
downloads

Cite This

Obtaining flow fields around airfoils is an essential process in aerodynamics design. With advanced technologies, simulating the whole process for any given mesh became possible using numerical analysis rather than wind tunnel studies with scaled models. However, computing high-accuracy solutions for flow fields requires considerably more powerful hardware and computation time. Deep learning models have shown outstanding performances in many fields in recent years, providing satisfactory predictions with much less inference time. Therefore, neural networks are a good alternative for simulations in situations where less accurate but quick solutions are needed, such as airfoil design. However, popular architectures like CNNs or RNNs are naturally unsuitable for those tasks. Hence, this thesis aims to utilize graph neural networks to predict flow fields around airfoils. Various airfoils with different angles of attack in near transonic regimes have been used as the dataset. Three studies have been done regarding graph neural networks in this thesis: (1) Different graph neural network architectures have been compared using the abovementioned dataset, (2) the importance of the surface nodes in training has been analyzed using different weights in the loss function, and (3) a proposed mesh-independent adaptive mesh refinement has been integrated into a deep learning methodology and node and area-based loss calculations are compared. Our studies found that graph neural network architectures perform similarly in this task, and surface nodes do not significantly affect prediction performance except for the pressure field. In the third study, we showed that the number of nodes in high-gradient areas can be increased in the same number of epochs without affecting the model's performance significantly. Besides, we pointed out that area-based loss and metric calculations are better in the representation of the distribution of flow field predictions.

Subject Keywords

Deep learning, Graph neural networks, Airfoil aerodynamics, Adaptive mesh refinement

URI

https://hdl.handle.net/11511/111388

Collections

Graduate School of Natural and Applied Sciences, Thesis