Comparison of Gradient-based and Global Autoencoded Lift-constrained Drag Minimizations

In this paper a traditional gradient-based quasi-Newton BFGS Optimizer as well as a global optimizer are applied to lift-constrained drag minimizations for steady inviscid flow over a transonic NACA 0012 airfoil with six Hicks-Henne sine bump shape design variables. For the global optimization, an autoencoder is used to perform a design dimensionality reduction and a surrogate model is constructed in the reduced dimension which is then employed in the global optimization. Results show that both methods lead to successful optimizations at comparable computational costs.