11:00   Data driven methods 2 Chair: Cedric Liauzun 11:00 30 mins Surrogate Modeling of Highly Flexible Structures and Aerodynamics Using Neural Networks Vitor Borges Santos, Breno Soares da Costa Vieira, Flávio Luiz Cardoso-Ribeiro, Antônio Bernardo Guimarães Neto Abstract: The renaissance of neural networks in the scientific community in recent years has brought new perspectives for improving the computational efficiency of traditional modeling techniques. Hamiltonian neural networks leverage the energy-preserving properties of the Hamiltonian formalism to provide surrogate models with increased interpretability compared to conventional feed-forward models. In this study, we employ a lumped-mass multibody method to derive the equations of motion of two highly flexible structures. We perform a model order reduction via modal decomposition while preserving the nonlinearities with the use of exact kinematic relations. After validating full- and reduced-order models, we use them to produce datasets and train the neural networks, which serve as ready-to-use surrogate models. Preliminary findings show that the surrogate models based on neural networks can significantly reduce the time necessary to simulate the free response of the structures. Furthermore, we demonstrate that surrogate models based on Hamiltonian neural networks have energy-preserving capabilities, maintaining accuracy levels even for long simulations. Due to their architecture, when external loads are considered, the surrogate models require the analytical calculation of the generalized forces, jeopardizing the efficiency gains obtained by our approach. We also present initial findings on the use of neural networks for faster aerodynamic models for flexible aircraft, particularly as surrogate models for the vortex-lattice method. By using a neural network as the aerodynamic surrogate model in a specific flexible aircraft simulation framework, the computational costs were reduced by a factor of 100, on average. The outcomes of this study demonstrate that surrogate models based on neural networks can soon become an efficient and reliable alternative for modeling arbitrarily flexible aircraft, provided the current limitations are addressed. 11:30 30 mins Nonlinear aeroelastic analysis of a regional airliner wing via a neural-network based reduced order model for aerodynamics Nicolò Laureti, Luca Pustina, Marco Pizzoli, Francesco Saltari, Franco Mastroddi Abstract: This paper investigates the use of a neural network based reduced order model to solve the nonlinear gust response of a regional aircraft wing characterised by high angle of attack and gust intensity. Dynamic aerodynamic stall is generally able to provide a natural load mitigation, which is generally not considered in aircraft design. The neural network is combined with strip theory, thus requiring the training of a single airfoil model. Given the difficulties in obtaining CFD based aerodynamic data due to the excessive time consuming simulations, the approach has started the training phase with data obtained from the Beddoes-Leishman unsteady aerodynamic model. The wing response is then evaluated by a reduced-order model implemented in the Simulink environment, based on Nastran structural data and strip theory, enhanced by a number of neural networks in parallel describing the nonlinear transient behavior. 12:00 30 mins Machine learning to forecast unsteady, motion-induced unsteady forces Reik Thormann, Hans Martin Bleecke Abstract: During flutter analyses of aircraft, a large parameter space has to be analyzed covering the extended flight envelope for a large set of mass variations and failure cases. With increasing aspect ratio and elasticity of the wing, flutter may become sensible with respect to the linearization point (lift coefficient, static aeroelastic equilibrium per dynamic pressure). Thus, assessing a good prediction of the flutter speed is crucial in a multi-disciplinary optimization already within the design phase. Therefore, machine learning techniques are explored to predict unsteady, motion induced unsteady forces for flutter analyses in a first step and secondly, consider amplitude-nonlinearities to support limit cycle analyses. The ROM setup is chosen such, that first the dimensionality of the problem is reduced. While for the structural deformations a transformation to modal coordinates is common, a so-called auto-encoder is used for the aerodynamic data. The modal coordinates are linked to the aerodynamic, latent subspace either via a Recurrent Neural Network (RNN) such as LSTM or a Transformer.

end %-->