Welcome to the GitHub repository for my 2023 Autumn semester project at EPFL CVLab! Here you can find the code, data, and report for my semester project. If you have any questions, feel free to email xingyue.zhang@epfl.ch.
This study explores the innovative application of visual prompting via image inpainting in the field of aerodynamic shape optimization, particularly for predicting pressure fields around airfoils. Our approach utilizes a 2x2 grid structure comprising airfoil images and their corresponding pressure fields. This grid structure is then fed into a pre-trained universal visual model that can predict the pressure field of an unseen airfoil based on the provided example, without any task-specific fine-tuning or model modification. We conducted extensive experiments to understand the effects of various factors such as airfoil size, color schemes of the pressure field, aggregation, and the similarity of airfoils used in the visual prompts. The results indicate that larger airfoils and specific color schemes enhance the model's accuracy. Aggregation also helps improve model performance. Moreover, using similar airfoils as examples significantly improves the performance compared to dissimilar ones.
The datasets we need are available at https://drive.google.com/file/d/1xZvFEghGXrVrVN5sSFdSvcCZm_sySjxW/view?usp=sharing. We only need the mesh and restart_csv folders from this link. Check UIUC Airfoil Data Site to learn more about the datasets.
The main code for this work is included in the airfoil_pressure_field.ipynb file. To run the model, use the model.ipynb file (credit to https://yossigandelsman.github.io/visual_prompt/).