Web-wing - Interactive Transonic Wing Design App

The physics-embedded transfer learning for transonic wing is demonstrated with a simple interactive app. The data-driven model is trained on RANS data, therefore has much better accuracy than existing fast aerodynamic prediction tools. Feel free to modify the airfoil geometry, wing planform geometry, and wing operating conditions to see what will happed on the wing surface flow field, and your knowledge on wing aerodynamics will grow.

The next step of the app is a gradient optimization tool for wing performance, which will come soon.

Installation

The app only has local-running version now because we can not afford a server. Please follow the steps below to run the code.

Install required packages

The web-Wing requires numpy, pytorch, matplotlib, and tqdm to be installed. The user interface is implemented with flask. You can install them with pip or conda.
Download libraries

The web-Wing requires three projects: floGen, cfdpost, and cst-modeling3d on GitHub. You can download the latest version from repositories, and unzip them. Intall three libraries with running python setup.py install at the root directory of each library.
Download trained model parameters

The model parameters can be download from the Tsinghua University cloud. Unzip the download and move the folder saves to <floGen folder>/examples/wing

Then, the wing-app should be able to start.

Start local server

change directory to <floGen folder>/examples/wing and run python app.py.
Start browser

go to 127.0.0.1:5000 in your browser.

User guide

The web-Wing is trained with RANS data and focus on the single section wing. The sectional airfoil is fixed along spanwise except its thickness.

The parameters include:

Parameter	Symbol	Definition	Range	Comments
CSTU	\(u_i, i=0,\cdots,9\)			These parameters defines the shape of the upper airfoil surface curve. Definition of CST parameters can be found in `cst-modeling3d`
CSTL	\(l_i, i=0,\cdots,9\)			These parameters defines the shape of the lower airfoil surface curve.
Root thickness	\( (t/c)_\text{max}\)		0.08 - 0.13	relative thickness of root
sweep angle	\(\Lambda_\text{LE}\)		0° - 35°	leading edge
dihedral angle	\(\Gamma_\text{LE}\)		0° - 3°	leading edge
aspect ratio	\(AR\)	\(\frac{b_{1/2}^2}{2S_{1/2}}\)	6 - 10
tapper ratio	\(TR\)	\(\frac{c_\text{tip}}{c_\text{root}}\)	0.2 - 1.0
tip twist angle	\(\alpha_\text{twist}\)		0° - -6°
tip-to-root thickness ratio	\(r_t\)		0.8 - 1
AOA	\(\alpha\)		1° - 6°	due to the limited training sample, the results under large AOAs may have large error
Mach	\(Ma\)		0.72 - 0.85

These values can be input with text box and scroll bars. You can also select the training wing parameters with select from droplist.

How it works

The model used here is the 2d-to-3d transfer model. We first build a model to predict 2d airfoil flow field, and then use it as a prior for 3d flow field prediction.

The UI is implemented with Flask.

To cite this work

Yang Y, Li R, Zhang Y, Lu L, Chen H. Rapid aerodynamic prediction of swept wings via physics-embedded transfer learning, submitted to AIAA Journal. https://arxiv.org/abs/2409.12711.
Yang Y, Li R, Zhang Y, Lu L, Chen H. Transferable machine learning model for the aerodynamic prediction of swept wings, Physics of Fluids, 2024. https://doi.org/10.1063/5.0213830.