Increased-Order Model for Unsteady Aerodynamic Nonlinearities Using Neural Networks

Discrete gust loads are computed based on linearized aerodynamics. Due to the large gust amplitudes, aerodynamic nonlinearities such as shock motion or flow separation may arise and cause significantly different loads compared to the linear prediction. Computational fluid dynamics (CFD) can be applied to account for these nonlinearities, as it has been demonstrated over the past decade for gust loads. Even though CFD is computationally expensive, the results can be used for data-driven surrogate models, which are fast to evaluate. This work suggests increasing the order of an aerodynamic state-space model by adding a nonlinear correction model to predict the nonlinear deviation from the linear solution. Here, a recurrent long short-term memory neural network can be applied. The nonlinear data stem from time-marching gust encounter simulations for the free-flying elastic aircraft at trimmed condition at transonic flow. The motion trajectories together with the gust are then forced into the linear model to calculate the nonlinear residual. The linear aerodynamic loads serve as an input to the neural network to predict the nonlinear load residual. The proposed model is demonstrated using the NASA Common Research Model configuration. The predicted aerodynamic loads are sufficiently close to the nonlinear reference simulations.