PINNPC: Physics-Informed Neural Network Predictive Control for MAV Transition Flight

Traditional neural networks are widely used for surrogate modeling of complex dynamics. However, they often suffer from poor interpretability, high computational costs, and an inability to provide the Jacobian matrices essential for model-based control. Physics-informed neural networks (PINNs), which embed physical laws into the training process, offer a promising alternative, yet their application to micro air vehicle (MAV) dynamics remains largely unexplored. This study proposes a PINN-based predictive control (PINNPC) framework for the transition flight of a vectored-thrust, tailless MAV. First, an analytical dynamics model incorporating propeller slipstream effects and multi-body interactions is established. A PINN is then trained as a high-fidelity surrogate of this physics-based model. Crucially, through automatic differentiation, the PINN directly supplies the Jacobian required for gradient-based optimization, enabling its seamless integration with nonlinear model predictive control. The proposed PINNPC framework achieves transition trajectory tracking in simulations, and demonstrates robustness under disturbances. This work presents the first application of a PINN for the transition control of a vectored-thrust tailless MAV, highlighting its potential to enhance model-based control performance.