Physics-based deep learning model for high-altitude stall recovery guidance

A deep learning (DL) based model is developed to predict flight trajectories, under highly nonlinear phenomena such as aerodynamic stall, and subsequent recovery attempts. Recurrent neural network is used as the fundamental architecture for the DL model due to the temporal dependencies of flight trajectories. The DL model acts as a reduced order model (ROM) and provides predictions of aircraft flight response based on the deflection profile of aircraft control surfaces over flight durations of interest. The NASA transport class model (TCM), which is a high-fidelity flight dynamics simulation framework, is used to simulate high-altitude stall upset conditions and subsequent flight trajectory recovery. Monte Carlo simulations coupled with TCM are used to generate a stochastic database of stall flight trajectories to account for uncertainties in the recovery process. The developed database is then used as the ground truth for the DL model training and validation. Upon successful reduced order model optimization, flight trajectories are computed without the need of TCM’s complex nonlinear aerodynamics and engine models, resulting in order of magnitude enhancement in computational efficiency. The developed methodology can be used for real-time flight trajectory predictions under aircraft upset conditions, which is a vital safety metric in enhancing the national airspace system.