Inverse Machine Learning Prediction for Optimal Tilt-Wing eVTOL Takeoff Trajectory

Abstract

The attention to electric vertical takeoff and landing (eVTOL) aircraft has grown significantly. Due to flexible maneuverability, eVTOL is suitable for missions required precise control, such as urban air mobility (UAM). Especially with the advances in battery technology, autonomous control, and ride-hailing services, large-scale UAM becomes achievable. To make UAM cost-effective and environmentally friendly, optimal trajectory design optimization is essential. However, high-fidelity simulation model-based trajectory design is computationally intensive due to iterative model evaluations. Therefore, this present work develops machine learning-based inverse mapping of optimal takeoff trajectory for a tilt-wing eVTOL. In particular, we adopt design requirement bounds (such as maximum allowable acceleration) as input and optimal trajectory design as output. We leverage long short-term memory (LSTM) for predicting optimal time sequential takeoff profiles and showcase resulting predictive performance against multi-output Gaussian processes (MOGP) surrogate. Results revealed that machine learning surrogates could reduce the computational time from 15∼20 minutes to several seconds on eight computing cores for each trajectory design case. Robustness test indicated that LSTM consistently outperformed MOGP by achieving over 99% relative accuracy predicting the optimal takeoff trajectory profiles.