Generative AI meets flight trajectories: a semantic trajectory representation approach

Recent advancements in Artificial Intelligence (AI) have opened new opportunities to enhance Air Traffic Management (ATM) performance. For ATM-oriented generative AI systems, understanding flight trajectory data is critical, as trajectories constitute the primary medium of information in air traffic operations. However, generative AI struggles with raw trajectory data because it is continuous, context-dependent, and spatially relative—unlike text or image tokens, which are discrete and semantically meaningful. Without specialized encoding or modeling, AI cannot effectively capture the structure or dynamics of trajectories. To address this challenge, this study proposes a semantic trajectory representation method to enable generative AI to better interpret flight trajectories. First, horizontal trajectories are decomposed into turning and straight segments, with a turning-point detection method isolating salient turning behaviors. Second, the decomposition is extended with a holding-pattern detection method to identify holding behaviors. Finally, trajectories are represented as sentences rather than arrays, leveraging the semantic structure extracted from the decomposition. A case study demonstrates that, compared with raw trajectory data, existing generative AI models achieve significantly improved trajectory understanding when using the proposed semantic representation.