M. Winter

For aircraft design and analysis, the accurate prediction of the static and dynamic interactions between the structural-elastic, inertial, and aerodynamic forces is of particular importance. In the context of those aeroelastic investigations, the highly-efficient linear potential flow methods have been mainly applied to model the time-varying aerodynamic forces induced by structural vibrations or gusts. However, the aforementioned methods do not fully meet todays accuracy requirements, especially in terms of transonic flow problems that are governed by intrinsic nonlinear effects. In contrast, the unsteady aerodynamic forces can be determined with sufficient fidelity using computational fluid dynamics (CFD) solvers. Since the latter methods require extensive computing resources, their industrial use for multidisciplinary analyses is still limited. Motivated by this efficiency bottleneck, model-order reduction methods based on machine learning approaches have been developed in this work to efficiently reproduce the unsteady aerodynamic characteristics. The time-domain reduced-order models (ROMs) are constructed by means of nonlinear system identification and neural network techniques, which allow the modeling and prediction of nonlinear flow phenomena. Hence, the aerodynamic ROM - represented in this work by a recurrent neuro-fuzzy model - reflects the essential dynamic relations of the underlying CFD system in a resource-saving manner and is capable, inter alia, of modeling the unsteady aerodynamic behavior subject to structural vibrations, varying freestream conditions, and pronounced shock motions. Based on the developed methodologies, motion-induced forces and moments or, in combination with the proper orthogonal decomposition (POD), locally-distributed aerodynamic loads can be computed in an accurate and robust way. Consequently, a significant efficiency enhancement of the CFD-based numerical analysis is achieved, which enables the investigation of the aircraft aeroelastic behavior at an earlier stage in the development process. The proposed approaches have been comprehensively tested and validated by test cases of variable complexity. In this regard, the computational efficiency as well as the fidelity of the ROMs has been assessed relative to the full-order CFD simulation procedure.

Deutscher Luft- und Raumfahrtkongress 2021

Citation

englisch

Machine Learning, Instationäre Aerodynamik, Data Science, Modelle reduzierter Ordnung, ROMs, Aeroelastik, Neuronale Netzwerke, Neuro-Fuzzy Modelle, Systemidentifikation, Nichtlineare Dynamik, CFD

Universitätsbibliothek der Technischen Universität München, München, 2021

13.10.2021