Numerical Validation of a Machine Learning–Driven Framework for Nonlinear UAV Flight Dynamics

Accurate numerical modeling of nonlinear flight dynamics is essential for the simulation, control, and performance assessment of unmanned aerial vehicles (UAVs). Classical numerical solvers such as the fourth order Runge - Kutta (RK4) method can provide high accuracy, but they often require small integration step sizes in strongly nonlinear operating regimes, which leads to increased computational cost. In this work, a physics informed, learning based correction framework is used to enhance the numerical solution of nonlinear UAV flight dynamics. The framework augments a baseline numerical solver with a neural correction module inspired by physics informed neural networks, so that the numerical solution is adjusted while remaining consistent with the governing dynamicsNumerical validation is carried out using flight operation data from 100 UAVs, including both fixed wing and multirotor configurations with different mission profiles. The data were obtained through collaboration with an industrial aviation partner and were anonymized and standardized before analysis to preserve confidentiality. The proposed framework is benchmarked against an RK4 solver taken as the reference. The results show that the learning-based framework reduces computational cost by about 38% while maintaining an RMSE of 1.82×10⁻³ rad in the pitch angle response. These findings indicate that the approach can provide accurate and efficient numerical solutions for nonlinear UAV flight dynamics and is suitable for real time and autonomous flight applications.