LLMs Generate Traceable Aviation Hazard Scenarios.

Operational safety analysis in aviation systems is complex, requiring consideration of numerous interacting factors such as weather, air traffic control actions, airspace constraints, aircraft operations, and human factors. This differs significantly from the functional hazard assessments applied at the aircraft-system level. This research introduces an AI-assisted approach designed to generate candidate hazard scenarios specifically for operational safety analysis, drawing data from NASA's Aviation Safety Reporting System (ASRS). The method takes a target adverse outcome and produces a structured hypothesis, detailing categorical factors, alongside a narrative scenario describing an operational event sequence consistent with that structure. Each generated scenario is accompanied by a plausibility score, derived from historical co-occurrence evidence, and crucially, traceability links to the most similar held-out ASRS reports. The paper also proposes a hybrid variant where narrative generation is conditioned on a structured hypothesis produced through evolutionary abduction. This variant demonstrates improved correctness and reduced variability in the generated scenarios. The study evaluates various large language models, prompting techniques (zero-shot vs. few-shot), and fine-tuning options to assess their impact on the validity and realism of both the structured hypotheses and the narrative scenarios.