Development of a CFD data-driven surrogate model using the neural network approach for prediction of aircraft performance characteristics

Development of a CFD data-driven surrogate model using the neural network approach for prediction of aircraft performance characteristics

Bourier, Sébastien (TU Delft Aerospace Engineering)

Hulshoff, S.J. (mentor)
van Rooij, Michel (graduation committee)

Delft University of Technology

Aerospace Engineering

2021-11-26

Evaluation of aircraft performance for design, certification, and maintenance purposes requires aerodynamic knowledge for the entire flight envelope of an aircraft. Simplified models that relate geometric properties and flight conditions of an aircraft to its aerodynamic properties are simply not sufficient anymore, as non-linear aerodynamic phenomena, e.g., vortex development and flow separation, can drastically influence the performance of an aircraft. Therefore, high-fidelity simulations and wind tunnel experiments are necessary to assess the performance of an aircraft sufficiently. Although the development of new technologies in the last centuries allowed the computational time of the high-fidelity simulations to be decreased significantly, the computational expense still remains large. New techniques that are used to simulate the load cases of an
aircraft and evaluate its aerodynamic properties are therefore aimed to provide an accurate representation of the high-fidelity simulations while the numerical complexity and hence the computational cost of the model is reduced. These techniques can be referred to as Reduced-OrderModels (ROM).

Previous studies, conducted at the Royal Netherlands Aerospace Centre (NLR) as part of a collaborative research task group within the Science and Technology Organization (STO) of the North Atlantic Treaty Organisation (NATO), investigated the development of different ROMs that should be able to accurately predict the surface pressure distribution of the MULti-DIsciplinary CONfiguration (MULDICON) Unmanned Combat Air Vehicle (UCAV). In particular, a ROM based on the neural network approach shows promising results but suffers from drawbacks such as inaccurate prediction of the surface pressure near the wing tip region and the exclusion of time-history effects. This thesis report will serve as a continuation of previous research done on the reduced-order modeling approach for the prediction of the surface pressure coefficients of the UCAV configuration and will provide a baseline for future research. The research objective for this thesis report is therefore to develop a CFD data-driven Reduced-Order Model (ROM) based on the neural network approach that is able to predict the surface pressure and integral aerodynamic load coefficients of the UCAV MULDICON design and is able to capture the transient effects of the flow field for different flight maneuvers with the capability to reach the same level of accuracy as high-fidelity tools.

In this thesis, a ROM is developed that makes use of a reduced-order basis that is constructed using Proper Orthogonal Decomposition (POD). The POD method decomposes the high-fidelity samples used to train the ROM into spatial POD modes that will be used for the reduced-order basis, that are ranked in order of contribution towards the total kinetic energy (TKE). A reduced-order basis with only 10 spatial POD modes already captures 99 % of the TKE, which includes the most dominant flow properties and can be used to project the full-order solution with an accuracy that is in the same orders of magnitude while the data that is used is minimized. A Recurrent Neural Network (RNN) architecture is used as a surrogate model to determine the input-output relationship of the model. For this thesis, a Long Short-Term Memory (LSTM) architecture is used, which is widely used for the prediction of time-sequential data. The ROM that is constructed in this thesis is a combination between the POD method and the LSTM architecture, which can be referred to as the
POD-LSTM model.

The performance of the POD-LSTM model has been evaluated and is split up into two different phases: the offline and online stages. The offline stage of the ROM is the stage where the reduced-order basis is constructed and the high-fidelity samples that are used for training are evaluated. From the offline stage, it can
be concluded that the high-fidelity samples provide good coverage of the regressor space and most of the test samples have the same projection error as the high-fidelity training samples that are used for the construction of the reduced-order basis. The POD-LSTM model has been trained for varying model parameters and it was shown that the number of LSTM units has the largest influence on the computational training time. In the online stage, the performance of the prediction of the surface pressure coefficients using test samples that are gathered from steady and unsteady harmonic pitch and plunge oscillations are evaluated. The results for the steady simulations show that the POD-LSTM model is able to accurately predict the axial and normal force coefficient, while inaccuracy is shown for the pitching moment coefficient. The main reason for the inaccurate prediction of the pitching moment coefficient is the inaccurate prediction of the surface pressure coefficient near the wing tip region. Both the harmonic pitch and plunge oscillations show similar results,
whereas the normal and axial force coefficients are predicted with good accuracy but the pitching moment coefficient is inaccurate due to an error in the surface pressure prediction near the wing tip region. The performance of the POD-LSTM model in terms of computational cost has been evaluated and was proven to be
faster than previous studies, whereas the number of LSTM layers and LSTM units has the largest influence on the computational performance.

Overall, the POD-LSTM model has provided solutions to improve on previous studies for the surface pressure coefficient prediction of the UCAV configuration. However, it must be noted that this thesis report serves as a baseline for future research and the full capabilities of the POD-LSTM model still need to be exploited to provide a fair comparison with previous studies. Different training maneuvers, increase in training time and different model architectures could influence the performance of the POD-LSTM model and therefore should
be incorporated in future studies to provide improvements to the proposed ROM.

CFD
Neural Network
Reduced-order model
POD
LSTM

http://resolver.tudelft.nl/uuid:a032f522-2d50-42ac-a50a-e3481bd484a9

Student theses

master thesis

© 2021 Sébastien Bourier