Key technologies and prospects of large models for decision support in aircraft carrier flight deck aviation operations

Aircraft carrier flight deck aviation operations constitute a critical process for generating and sustaining the sortie generation capability of carrier-based aircraft. Their high-risk, highly constrained, and tightly coupled nature places stringent demands on the safety, reliability, and interpretability of intelligent decision-support systems.

Focusing on decision support for aircraft carrier flight deck aviation operations, this paper systematically reviews the foundational research on rule-based constraints, situational awareness, and scheduling optimization. It further examines the key challenges associated with the application of large models in this domain, including trustworthy output, multimodal cross-domain fusion, adaptation to diverse operational scenarios, and causal reasoning across multiple operational processes. Using two representative scenarios as case studies—pre-launch coordinated support involving aircraft towing, weapon loading, refueling, and catapult-window coordination, and dynamic replanning during recovery operations under conditions such as arresting-gear malfunctions or parking-space congestion—this paper analyzes task objectives, input modalities, operational constraints, output forms, validation mechanisms, and performance evaluation metrics. Based on this analysis, the paper clarifies how large models can be integrated into practical support workflows. The former scenario emphasizes resource coordination under efficiency and safety constraints, whereas the latter focuses on risk mitigation and recovery-oriented scheduling in the presence of operational disturbances. Together, these scenarios encompass both routine operational planning and contingency-driven reconstruction in flight deck aviation support. On this basis, this paper proposes a large-model-driven decision support framework for high-safety-level aviation support operations. The framework establishes a closed-loop decision-making process integrating generation, verification, screening, and confirmation. First, physical mechanisms, operational regulations, spatial constraints, safety separation requirements, and process dependencies are incorporated into the model generation process. In parallel, external verification is achieved through the integration of rule engines, collision detection, formal verification, discrete-event simulation, and digital-twin-based reasoning. Second, to handle heterogeneous information sources, such as flight plans, deck surveillance videos, target-position time series, voice commands, equipment status data, engineering geometric constraints, and operational logs, a unified multimodal representation and encoding framework is developed to support situational awareness, event correlation, and task-state representation. Third, retrieval-augmented generation, lightweight fine-tuning, prompt templates, and scenario recognition are integrated to establish a domain-knowledge-driven cross-scenario adaptation mechanism, thereby enhancing the model's transfer capabilities across tasks such as launch, recovery, refueling, weapon loading, maintenance, and contingency response. Finally, for continuous operational processes including towing, support, launch, recovery, and maintenance, a process-level causal reasoning and multi-agent collaboration framework is constructed to characterize how local disturbances propagate through downstream tasks, trigger resource reallocation, and drive process reconstruction. Furthermore, considering shipborne edge-computing conditions, an engineering deployment framework is proposed that integrates lightweight inference, local knowledge enhancement, rule- and simulation-based verification, and command workflow integration. Structured outputs are adopted to represent task objects, state evidence, candidate actions, verification outcomes, risk levels, and human-confirmation status.

The results show that large models for flight deck aviation support should not be employed as direct generators of execution-commands. Instead, they should function as decision-support modules for candidate-scheme generation, situational interpretation, risk assessment, and process-level reasoning within a closed-loop framework incorporating local verification and human confirmation.

The proposed framework can effectively mitigate untrustworthy outputs, while enhancing multimodal situational awareness, cross-scenario adaptability, proactive replanning capability, and engineering integration capability. It provides both theoretical foundations and methodological guidance for the safe and controllable application of large models in complex military support systems.