Self-Supervised Airfoil Shape Optimization

Aerodynamic Shape Optimization (ASO) aims to enhance aerodynamic performance by modifying a body's shape using computational fluid dynamics and adjoint methods. Though effective, these methods are computationally intensive. Surrogate-Based Optimization (SBO) via machine learning offers a faster alternative but struggles with high-dimensional problems due to the exponential increase in required samples. To address these limitations, this paper explores Self-Supervised Learning (SSL) for ASO. SSL directly predicts optimal shape variables from problem-instance parameters, avoiding the curse of dimensionality associated with traditional methods. A simple airfoil example is used to demonstrate SSL's potential, showing it generalizes better than SBO and other conventional neural network approaches. This suggests SSL as a promising scalable solution for expansive aerodynamic optimization tasks, but future work remains to verify scalability on an industrial size aerodynamic problem.