Wings in a Minute: How AI Became an Aerodynamic Genius ⚡ экспресс
A methodology for efficiently constructing accurate surrogate models for aerodynamic computations in shape optimization tasks is proposed. The AeroTransformer model, based on the Transformer architecture, is first pre-trained on the SuperWing dataset containing nearly 30,000 samples of diverse wing geometries, and then fine-tuned on a small number of target samples for a specific shape. In experiments with transonic wings, where the shape varied around the Common Research Model, using 450 target samples achieves a surface flow prediction error of 0.36%, which is 84.2% lower than training from scratch. A systematic analysis of the impact of architectural choices and training strategies for limited data and computational resource settings was also conducted. For reproducibility, datasets and pre-trained models are published, along with an interactive design tool.
Training an accurate aerodynamic model is like culinary art: first, a chef masters hundreds of recipes to grasp the general principles, then easily perfects a specific dish. Engineers created AeroTransformer and fed it 30,000 different wing shapes — a whole galaxy of culinary ideas galaxy. In calculations, air behaves like water in a pot water: smoothly flowing around surfaces, simplifying computer prediction.
After fine-tuning on just 450 samples, error dropped to 0.36% — six times more accurate than starting from scratch. The approach sharply reduces entropy entropy (uncertainty) in predictions.
An online tool already lets you change wing shape in a browser and see the result instantly. The developers have open-sourced the code and data — now any researcher can fine-tune the model for turbines or racing cars.
🎯 The model is already working on the WebWing website: move the wing with your mouse and watch lift in real time.