Manifold Learning of Nonlinear Airfoil Aerodynamics with Dimensionality Reduction

This paper aims to explore the advantages offered by machine learning (ML) for dimensionality reduction of nonlinear transonic aerodynamics. Three ML techniques are evaluated in terms of their ability to generate interpretable low-dimensional manifolds of the transient pressure distributions over a NACA4412 airfoil equipped with a flap. These ML techniques are Kernel Principle Component Analysis (kPCA), Locally Linear Embedding (LLE), and t-distributed Stochastic Neighbourhood Embedding (t-SNE). Initial investigations are also carried out to evaluate the performance of Artificial Neural Networks (ANNs). Three transient aerodynamic test cases are evaluated. First, a static aerodynamic transient analysis. Second, pitching and heaving airfoils in terms of prescribed sinusoidal displacements. Lastly, the airfoil geometry is adapted to include a flap under sinusoidal actuation. The snapshots forming the ground truth are obtained from unsteady CFD simulations. The preliminary results of this study reveal that patterns exist in low-dimensional nonlinear manifolds. Furthermore, unsupervised learning techniques are seen to outperform supervised neural networks in terms of both training cost and reconstruction accuracy. Promising reconstruction capabilities are observed with unsupervised learning.