Flight Dynamics Modeling Using Physics-Informed Neural Networks

In this work, physics-informed neural networks (PINNs) were implemented as a parameter estimation and system identification technique that can be used on flight data. PINNs were, in part, developed as a parameter estimation technique in which physical knowledge of a system is directly introduced into the neural network training. Variations of the generalized PINN structure were developed to be implemented in a framework for experimental flight testing. The first variant, Determinant PINNs, consisted of a trajectory network and a parameter estimation module, which can be used to estimate parameters that vary due to changes in explanatory (independent) variables, such as angle of attack within a determinant (predetermined) form. Second, Non-Determinant PINNs consisting of a trajectory network and a parameter estimation network can be used to estimate parameters that vary due to a set of explanatory variables without using a determinant form. Third, Modified Non-Determinant PINNs were developed to demonstrate the use of a nondeterminant parameter estimation network independent of a trajectory network while implementing physical constraints into the training process. These variations extend the parameter estimation abilities of generalized PINNs from constant parameters to varying parameters and system identification. Finally, multiple simulated case studies were used to show results that provide promise for further investigation in increasingly challenging regimes.