Session: 02-06-01: Machine Learning in Structural Dynamics and Aeroelasticity

Paper Number: 183357

183357 - Rapid Predictions of Nonlinear Flight Dynamics for Rigid Aircraft With Physics-Informed Neural Networks 

This paper introduces linked physics-informed neural networks (PINNs) as efficient surrogate models for the nonlinear dynamics of rigid-body aircraft. The dynamics of rigid-body aircraft are accurately described by nonlinear differential equations that continuously represent the changes in the vehicle's state over time, based on the loads and control inputs it experiences. However, directly integrating these nonlinear equations can become a bottleneck in various real-time applications. In this work, a modified PINN is formulated to provide rapid predictions of aircraft flight dynamics. This approach overcomes two significant limitations of traditional PINNs, namely: (1) PINNs are typically designed and trained on a single system, and must be retrained for each system of interest, even when the governing physics are identical, and (2) PINNs are trained on a bounded time duration, outside of which their accuracy significantly diminishes. These disadvantages hinder the adoption of traditional PINNs in time-sensitive applications, such as real-time feedback control and digital-twin simulations. This work outlines a technique to overcome these limitations by augmenting the PINN input with the initial state and time coordinate of the target and chaining the subsequent PINN solutions in an end-to-end manner. As a result, the modified PINN is trained to reproduce a system's response for various bounded initial conditions, eliminating the need to retrain for different bounded time intervals. This architecture offers a computationally efficient and accurate surrogate model, speeding up the simulation by preemptively addressing the computational demands of training. To facilitate simulating closed-loop controlled flight, the PINN is integrated with a simple linear-quadratic regulator (LQR) controller. The prediction accuracy for several basic maneuvers is evaluated by comparing it with traditional numerical integration using a Simulink-based numerical model of the same dynamics and identical control gain. This work demonstrates the feasibility and advantages of linked PINNs in fast dynamics simulation and control settings.

Presenting Author: Augustine Loshelder The University of Alabama

Presenting Author Biography: Augustine Loshelder is a PhD Candidate at the University of Alabama's Aeroelasticity and Structural Dynamics Research (ASDR) Lab, advised by Dr. Weihua Su. He has received multiple awards for his research including departmental awards for outstanding research and best-in-category paper at the SMASIS 2024 conference.