The development of realistic synthetic trajectory data is essential for the planning and validation of future Air Traffic Management concepts. Deep generative models have emerged as state-of-the-art methods for synthetic trajectory generation due to their ability to learn complex dynamics. In particular, Variational Autoencoders (VAEs) have gained significant attention because of their stable training. However, for the generated data to be trusted, model interpretability is crucial. A key aspect for interpretability is the analysis of the latent space, where the essential features and patterns of trajectory behavior are encoded. However, in practice, many latent dimensions remain weakly utilized or collapse toward the prior, limiting interpretability and effective use of latent capacity.

In this work, we study latent dimension activity in temporal VAEs trained on real-world aircraft trajectory data, with a specific focus on how Kullback-Leibler (KL) divergence regularization, a penalty added to the model’s loss function, affects the number of active latent dimensions. We adopt a temporal convolutional VAE with fixed latent dimensionality and formulate the KL divergence as an average over latent dimensions, thereby controlling the information content per dimension.

Additionally, we introduce a modified training objective that explicitly suppresses near-collapsed dimensions through masking and reweighting of the KL term. Our results show that this intervention significantly increases the number of active latent dimensions without degrading reconstruction accuracy. Ultimately, this will improve understanding of the decisions behind trajectory generation, enabling wider adoption. ... Read more