Utilizing Artificial Neural Networks for Entry Vehicle Aerodynamic Characterization

Abstract

Determining the dynamic stability of blunt body entry vehicles is a persistent engineering challenge, particularly in the low supersonic to subsonic flight regime where the behavior of the unsteady wake is a primary contributor. Dynamic stability quantities are determined by fitting measurements of a ballistic range campaign or a computational fluid dynamics (CFD) computational experiment to an assumed functional form in order to regress quasi-static stability coefficients. However, this data reduction process has many implicit assumptions that may not hold. This paper explores novel alternatives to the established methods for modeling blunt body aerodynamics. A six-degree-of-freedom CFD-in-the-loop flight model is used to run “virtual ballistic range tests,” fully capturing the relevant flow physics. Feed-forward and time-delay neural network models are fitted to the time-series trajectory and aerodynamic results, which can then be used to predict aerodynamic forces and moments. These models do not have a prescribed functional form and do not assume linearized aerodynamics. The models are evaluated for goodness-of-fit in their aerodynamic and trajectory prediction. The feed-forward neural network model resulted in a better prediction of the virtual ballistic range tests than a traditional database. The time-delay network had good open-loop performance but suffered from closed-loop instability.