Conditional Variational Autoencoders for aircraft type-specific trajectory generation

Trajectory generation has traditionally relied on methods based on physical aircraft performance models. Recently, data-driven machine learning approaches have demonstrated in- teresting statistical properties, proving useful in applications such as collision risk modelling and risk estimation, albeit often at the expense of physical realism. Physics-informed machine learning is a promising long-term solution to this limitation, but such models are typically difficult to train in practice. In this paper, we investigate an alternative approach by adapting a VampPrior- based Variational Autoencoder (VAE) architecture from a previous contribution into a Conditional Variational Autoencoder (CVAE). This enables the training of a single model that can handle multiple aircraft types. The proposed CVAE achieves performance comparable to that of dedicated VAE models trained separately for each aircraft type, while additionally addressing two challenges: (i) providing a generative model for sparsely represented aircraft types, for which no specific model can be trained, and (ii) improving overall regularisation of the generated trajectories.