New generation combat aircraft are expected to operate over extended flight envelopes, including flight at high flow angles and rapid maneuvers. Conditions beyond traditional limits are giving rise to nonlinear phenomena, such as flow separation, large scale energetic vortices, fluctuations etc. These phenomena have significant impact on aircraft performance and if not resolved accurately design uncertainties are increased risking lack of performance or even costly redesigns. Thus, accurate modelling of unsteady nonlinear aerodynamics is essential for modern and future combat aircraft.
Unfortunately, conventional modelling tools either lack the required fidelity or they are too expensive. Traditional, highly-efficient approaches are not suitable for modelling nonlinear flow phenomena. Concurrently, high fidelity Computational Fluid Dynamics (CFD) simulations are computationally demanding and therefore impractical in many cases. To enhance aircraft design, it is desirable to obtain models joining the best of these two worlds. A common approach is to distill high fidelity methods into Reduced-Order Models (ROMs) that can accurately approximate unsteady aerodynamics at orders of magnitudes lower costs than CFD. Relevant literature offers many different ROM techniques for varying purposes. Nonetheless, constructing such models is still challenging and currently there is no generally agreed method.
In the current thesis a ROM technique that may be applicable to wider ranges of problems and simpler to construct is sought. The objective is to obtain a model that can promote aircraft control design, performance assessment and structural analysis throughout dynamic maneuvers over complete flight envelopes. The thesis proposes a novel approach utilizing modern, deep convolutional neural networks (CNNs). The devised model consists of three main components. First it incorporates a geometry description constituted by coordinates of an aircraft CFD surface grid. Second, a primary encoding-decoding CNN predicts pressure distribution at the grid points of the geometry. The final and third part of the model is an auxiliary encoding CNN deriving integral aerodynamic loads corresponding to the pressure field predictions of the primary network. The model evaluates and produces instantaneous values. Given a maneuver, it proceeds in timesteps. The predictions of the separate instances are computed directly without the need of subiterations (as it would be the case for CFD simulations).
As a proof of concept, the model is applied to symmetric motions in the vertical plane at fixed Mach number and altitude. The subject of the investigations is the MULDICON configuration of the 251 th Science and Technology Organization work-group of NATO. To fully exploit the advantages of reduced-order modelling, flow characteristics are inferred from a single excitation following an efficient system identification technique using Schroeder sweeps as input signals. The performance of the model is assessed by numerous test cases performed in CFD. First, steady conditions of varying incidence angles are investigated. Second, harmonic pitch and plunge oscillations around different angles of attack at different amplitudes and frequencies are considered. Third, additional test cases of a linear pitch up-down -- and a climbing maneuver are studied.
Considering computational efficiency, the results show robust model performance. GPU-accelerated CNN calculations are conducted roughly 5000 times faster than CFD simulations. The primary network can accurately resolve the pressure distributions over large portions of the geometry. Lower surface predictions are very accurate. However, among certain conditions discrepancies are observable on the upper surface towards the wingtips. Still, the secondary network can predict corresponding aerodynamic forces accurately. In contrast, its moment predictions are sensitive to errors in pressure distributions. Consequently, moment predictions can largely deviate from reference data, especially when nonlinear phenomena are prominent. However, in many cases errors are attributed to insufficient regressor space coverage, i.e.\ certain input combinations are explored poorly by the Schroeder sweeps. Reconsidering system identification practices might mitigate those issues. Nevertheless, the thesis proves the applicability of deep CNNs to the problems at hand. Additionally, the results encourage further investigations.