Nonlinear Aeroelastic Reduced Order Modeling by Recurrent Neural Networks

The paper develops a reduction scheme based on the identification of continuous time recursive neural networks from input-output data obtained through high fidelity simulations of a nonlinear aerodynamic model at hand. The training of network synaptic weights is accomplished either with standard or automatic differentiation integration techniques. Particular emphasis is given to using such a reduced system in the determination of aeroelastic limit cycles. The related solutions are obtained with the adoption of two different approaches: one trivially producing a limit cycle through time marching simulations, and the other solving a periodic boundary value problem through a direct periodic time collocation with unknown period. The presented formulations are verified for a typical section and the BACT wing.