An Exploratory Assessment of LLMs’ Potential for Flight Trajectory Reconstruction Analysis

Large Language Models (LLMs) hold transformative potential for analyzing sequential data, offering an opportunity to enhance the aviation field’s data management and decision support systems. This study explores the capability of the LLaMA 3.1-8B model, an advanced open source LLM, for the tasks of reconstructing flight trajectories using synthetic Automatic Dependent Surveillance Broadcast (ADS-B) data characterized by noise, missing points, and data irregularities typical of real-world aviation scenarios. Comparative analyses against traditional approaches, such as the Kalman filter and the sequence to sequence (Seq2Seq) model with a Gated Recurrent Unit (GRU) architecture, revealed that the fine-tuned LLaMA model significantly outperforms these conventional methods in accurately estimating various trajectory patterns. A novel evaluation metric, containment accuracy, is proposed to simplify performance assessment and enhance interpretability by avoiding complex conversions between coordinate systems. Despite these promising outcomes, the study identifies notable limitations, particularly related to model hallucination outputs and token length constraints that restrict the model’s scalability to extended data sequences. Ultimately, this research underscores the substantial potential of LLMs to revolutionize flight trajectory reconstruction and their promising role in time series data processing, opening broader avenues for advanced applications throughout the aviation and transportation sectors.