Generative AI for Air Traffic Communication Automation - Fine-Tuned Language Models Using Simulation-Grounded and Real-Trace Data

#autonomous-systems #learning #planning #management #agents

Air Traffic Control (ATC) communication is a safety-critical process governed by strict procedural and phraseological standards, yet most prior applications of Large Language Models (LLMs) in aviation have focused on isolated tasks such as speech recognition or static regulatory question answering. This work proposes a three-stage adaptation pipeline for predictive, context-aware ATC communication. First, open-weight LLMs are domain-adapted on FAA/ICAO regulatory documents and aviation question-answer pairs to establish procedural alignment. Second, the models are instruction-tuned on structured ATC dialogue templates to learn sequencing, role consistency, and standardized phraseology. Third, the adapted models are re-tuned using simulation-grounded Digital Twin (DT) traces from Greensboro Airport (KGSO), enabling response generation conditioned on evolving aircraft state, runway configuration, and operational context. 
We evaluate open-weight LLMs across these stages using both semantic and operational metrics. 
Results show that high accuracy can be achieved, with Phi-4 producing the strongest overall performance and simulation grounding substantially improving context-aware ATC generation. These findings suggest that combining domain adaptation, phraseology-aware instruction tuning, and DT-guided re-tuning provides a promising foundation for simulation-driven controller training, decision support, and future human-in-the-loop ATC systems.