Since the publication of Tensorflow in 2015, neural networks have been widely adopted as a key tool for enabling applications ranging from image processing in self-driving cars, to drop-in replacements for costly simulations in design workflows. In recent work, Bharath Govindarajan and I explored the use of neural networks to accelerate airfoil and blade design workflows. We represent an airfoil with 10 design variables and used a Multi-Objective Optimization strategy to identify the best hover / cruise efficiency trade-off for a prop-rotor, including the effect of changing airfoil shapes in the design workflow.
Key take-aways:
Chebyshev-CST parameterization can represent > 1200 airfoils in the UIUC database with 10 design variables
Optimizing airfoils and blades separately can yield sub-optimal rotors, especially when airfoil design targets are geared towards classical "helicopter" style metrics
Hover-optimal airfoils are highly cambered; cruise-optimal airfoils have low camber and minimum drag at low lift coefficients
The neural network is two orders of magnitude faster than XFOIL, and 4-5 orders of magnitude faster than 2D CFD
Neural network surrogates decouple the computational complexity of optimization from that of the physics model used to generate training data