Nonlinear aerodynamic reduced order modeling by discrete time recurrent neural networks

Nowadays, viable estimations of transonic aerodynamic loads can be obtained through the tools of computational fluid dynamics. Nonetheless, even with the increasing available computer power, the cost of solving the related non-linear, large order models still impedes their widespread use in conceptual/preliminary aircraft design phases, whereas the related nonlinearities might critically affect design decisions. Therefore, it is of utmost importance to develop methods capable of providing adequately precise reduced order models, compressing large order aerodynamic systems within a highly reduced number of states. This work tackles such a problem through a discrete time recursive neural network formulation, identifying compact models through a training based on input-output data obtained from high-fidelity simulations of the aerodynamic problem alone. The soundness of such an approach is verified by first evaluating the aerodynamic loads resulting from the harmonic motion of an airfoil in transonic regime and then checking aeroelastic limit cycle oscillations inferred from such a reduced neural system against high fidelity response analyses.