PIGD-TL: Physics-Informed Generative Dynamics with Transfer Learning

This study investigates the fusion of physics-informed machine learning (PIML) and generative adversarial networks (GAN), targeting specifically the flight dynamics modeling of Electric Vertical Takeoff and Landing (eVTOL) aircraft. Emphasis is placed on the beneficial role of transfer learning, which utilizes pre-existing models for cognate tasks, leading to heightened model efficiency and decreased computational expenses. This approach is particularly advantageous in the realm of PIML, facilitating model applicability across similar systems or scenarios. The research notably introduces a novel set of models capable of generating physically viable scenarios, thereby increasing the authenticity and effectiveness of digital twins for system dynamics prediction. Validation of this methodology through a series of flight simulations and tests reveals a significant decrease in prediction error post-transfer learning implementation. The PIGD model, for instance, shows an average error reduction of 11.30% relative to the model trained using flight test data, thus enabling more accurate predictions with scarce flight data and mitigating issues related to extensive flight tests such as safety, time, and cost. This study considerably aids in the advancement of urban air mobility, especially concerning eVTOL aircraft, and sets the stage for future explorations and applications of PIML in intricate domains.