An unsteady aerodynamic reduced-order modelling framework for shock-dominated flow with application on shock-induced panel flutter prediction

A fully data-driven unsteady aerodynamic reduced-order modelling framework based on the nonlinear autoregressive with exogenous input structure is established for fluid-structure coupling simulations in the shock dominated flow. A generalized radial basis function neural network extended with polynomials is used for mapping the regressors and model outputs. The training with validation techniques is adopted to enhance the model's generalization ability, and the Bayesian optimization algorithm is selected for hyperparameter tunning. In addition, the fluid-structure coupling is incorporated into the validation process with a modified loss function to improve the robustness of the trained models. Both a generalized aerodynamic force model and a proper orthogonal decomposition based distributed aerodynamic force model are constructed and tested for the prescribed surface motions and fluid-structure coupling simulations. The results show that the constructed models have high accuracy in the forced oscillation tests, and the predicted amplitudes and frequencies of limit cycle oscillations in the shock-induced panel flutter are in excellent agreement with computational fluid dynamics/computational structural dynamics coupling simulations. The statistical results show that the online computational cost of the reduced-order model are orders of magnitude less than that required for the computational fluid dynamics method, indicating the presented modelling framework is an effective tool for the shock dominated aeroelastic problem analysis.