Aeroelasticity & Structural Dynamics in a Fast Changing World
17 – 21 June 2024, The Hague, The Netherlands
Home Program Author Index Search

Machine learning-based parametric model order reduction for the gust load analysis


Go-down ifasd2024 Tracking Number 135

Presentation:
Session: Reduced-order modelling
Room: Room 1.6
Session start: 09:40 Thu 20 Jun 2024

Sangmin Lee   sj7714@snu.ac.kr
Affifliation: Seoul National University

SiHun Lee   s.hun.lee@samsung.com
Affifliation: Samsung Electronics Company

Younggeun Park   pyk31213@snu.ac.kr
Affifliation: Seoul National University

Seung-Hoon Kang   shkang94@snu.ac.kr
Affifliation: Seoul National University

Kijoo Jang   prastins@samsung.com
Affifliation: Samsung Electronics Company

Haeseong Cho   hcho@jbnu.ac.kr
Affifliation: Jeonbuk National University

SangJoon Shin   ssjoon@snu.ac.kr
Affifliation: Seoul National University


Topics: - Reduced Order Modeling (High and low fidelity (un)coupled analysis methods:)

Abstract:

Gust load analysis plays an important role in the design process of an aircraft. The medium fidelity aeroelastic analysis such as that by doublet-lattice method (DLM) offer an efficient way of computing aerodynamic loads regarding the gust. Although DLM offers efficient aeroelastic analysis, significant amount of the computational time will still be required when multi-parametric study is required. As a substitute, parametric model order reduction (pMOR) will be considered to alleviate the computational time along with sufficient accuracy. Here, to approximate the nonlinear full order model (FOM) successfully with a slowly decaying Kolmogorov n-width problem, nonlinear pMOR method is considered. More specifically, nonlinear pMOR method adopting generative machine learning methods, variational autoencoder (VAE) will be considered. In this research, results by nonlinear pMOR with multiple parameters will be presented for the prediction of the gust response. The nonlinear pMOR and prediction will be attempted by LSH-VAE [2]. LSH-VAE is capable of accurate pMOR even when dealing with a large FOM dataset with an aid of the deep hierarchical structure. The deep hierarchical structure, composed of the bidirectional encoder-decoder group, will be adopted to mitigate the loss of long-range correlation. The loss function of LSH-VAE consists of a hybrod weighted least squares and Kullback-Leibler divergence loss to improve performance. Subsequently, in the reduced-dimensional latent space, the target latent variable will be acquired by conducting spherically linear interpolation. The decoder network will be used to generate the interpolated object from the interpolated latent vector. The current scheme will be applied for a high altitude long endurance aircraft under a discrete gust. To perform pMOR, the gust amplitude and duration will be considered as the parameters. The FOM will be obtained using a DLM-based analysis, ZAERO, for 10 baseline parameters selected by Latin hypercube sampling. For the accuracy evaluation, an unanalyzed parameter will be selected. Interpolated structural responses such as the wing tip displacement and internal stresses will be compared to the results by the unanalyzed parameter. Then, the computational efficiency will be evaluated. Finally, the structural response that satisfies certain range of the flight condition will be sought by using LSH-VAE.