Surrogate Unsteady Aerodynamic Modeling with Autoencoders and Long-Short Term Memory Networks

This paper presents preliminary results of an ongoing research in prediction of time-dependent flow fields by focusing on data-driven surrogate modeling using artificial neural networks for unsteady aerodynamic problems. The aim of this research is to model unsteady flow fields with learning in low-dimensional space and reconstruct with recurrent autoencoders. Within the scope of this paper, we separately share our findings in viscous unsteady flow field reconstruction of a 2D cylinder in a channel with a deep autoencoder and unsteady aerodynamic-acoustic time-series prediction of the supersonic NASA C25D aircraft with shallow long short-term memory networks. Satisfactory results are achieved in both unsteady applications, yet further improvements and validations are needed to be achieved to establish the desired surrogate unsteady aerodynamic modeling for supersonic aircraft maneuvers.