wchen459
commented
6 months ago Thanks for your interest. Except for trying to add a control point constraint as mentioned in https://github.com/IDEALLab/bezier-gan/issues/2#issuecomment-753245497, you may also want to increase your latent dimension. Based on our study (please refer to our AIAA journal paper), a latent dimension of 8 well captured the variance in the UIUC data. With fewer latent dimensions, there is the risk of making the discriminator out-compete the generator too much so that the generator cannot learn properly.
Hello,
First, I want to express my admiration for the incredible work you've done with this project.
Issue Description I've encountered an issue with the airfoil shapes generated by the model, specifically near the trailing edge. It appears that the lower curve occasionally crosses over the upper curve, which, as you might imagine, could impact the aerodynamic properties of the airfoil. I've attached an image below for your reference. My latent dimension is 3 and the noise one is 5.
Has this issue been observed during the development process? I noticed a similar topic discussed in Issue #2 and plan to attempt the solutions suggested there as an initial step towards mitigating this problem.
Beyond the initial mitigation steps, I'm considering implementing a post-processing script that identifies the crossing point and "cuts" the airfoil at that location. Subsequently, the script would re-normalize the airfoil coordinates to fit within the [0, 1] range across the chord. I don't really care for perfect representations of the original airfoils, but mostly for realistic airfoil with a low number of design variables.
Waiting for your thoughts.