Flight Dynamic Uncertainty Quantification Modeling Using Physics-Informed Neural Networks

When attempting to develop aerodynamic models for extreme flight conditions, including high angle of attack, high rotational rates, and tumbling motion, many classical methods have challenges in accurately modeling the highly nonlinear aerodynamic effects present. Physics-informed neural networks (PINNs) have previously been shown to be a potential technique to model these nonlinear aerodynamic effects when framed as a system identification problem. PINNs are well suited to this problem as they benefit from the universal approximation abilities of neural networks while directly incorporating known physical constraints into the training process. One of the main challenges in machine learning algorithms, including PINNs, is quantifying the confidence in a deterministic model prediction. This work expands on the previous development of PINNs as an aerodynamic and system identification tool by incorporating uncertainty quantification through three ensemble methods to provide calibrated confidence intervals on both aerodynamic coefficients and propagated trajectories. This work demonstrates and evaluates these methods on a simulated F16 case study where up to 100 PINN models are trained on varying training datasets. These models provide aerodynamic coefficients directly and are used to propagate trajectories within a six-degree-of-freedom simulation environment.