No abstract

Neural networks are applied to create reduced-order models (ROMs) for high-fidelity, nonlinear steady and unsteady CFD aerodynamic simulations by non-intrusively relating outputs (dependent variables) to inputs (independent variables), regardless of the complex physics involved. The present study is conducted with increasing complexity in the aerodynamic system and the corresponding neural network-based ROMs. The primary goal of this paper is to introduce the development and demonstration of new techniques for improving the predictive performance of the reduced-order models. In particular, the benefits of a physics-driven approach for selecting the training inputs and delay states are investigated. The adaptive ROM system is established by performing error estimation through two statistical techniques: cross-validation and bootstrapping.

No abstract

In this paper, the accuracy and practical capabilities of three different reduced-order models (ROMs) are explored: an enhanced implicit condensation and expansion (EnICE) model, a finite element beam model, and a finite volume beam model are compared for their capability to accurately predict the nonlinear structural response of geometrically nonlinear built-up wing structures. This work briefly outlines the different order reduction methods, highlighting the associated assumptions and computational effort. The ROMs are then used to calculate the wing deflection for different representative load cases and these results are compared with the global finite element model (GFEM) predictions when possible. Overall, the ROMs are found to be able to capture the nonlinear GFEM behaviour accurately, but differences are noticed at very large displacements and rotations due to local geometrical effects.