Artificial Intelligence-Driven Flight Disruption Management: A Scoping Review and Critical Evidence Synthesis for Predictive Decision Support in Airline Operations Control Centers

Technical Summary: Artificial Intelligence-Driven Flight Disruption Management

Problem Statement
Airline Operations Control Centers (OCCs) manage disruptions within a tightly coupled socio-technical system where weather, airspace constraints, aircraft availability, crew legality, and passenger connections interact. The core challenge is not merely predicting delays but managing the time-critical recovery of resources (aircraft, crew, passengers) under hard constraints and incomplete information. A critical gap exists between accurate disruption prediction and actionable disruption management: high predictive accuracy does not inherently translate to lower disruption costs, improved recovery feasibility, or better operational outcomes. Current literature often treats prediction and recovery as isolated problems, failing to address the integration required to convert forecasts into feasible, explainable, and uncertainty-aware recovery actions in live OCC environments.

Methodology
This paper is structured as a scoping review and critical evidence synthesis. It does not present new empirical experiments, simulations, or operational trials. Instead, it maps the existing heterogeneous literature across machine learning, operations research, complex networks, and human factors.

• Search Protocol: A reproducible search strategy is defined for major databases (Scopus, Web of Science, IEEE Xplore, etc.) using Boolean logic targeting flight disruption, AI/ML methods, and OCC decision support.
• Synthesis Framework: Evidence is synthesized across five research questions (RQs) covering: (1) AI/ML and optimization approaches; (2) data sources and validation strategies; (3) the linkage between prediction and recovery decisions; (4) barriers to deployment (explainability, safety, human factors); and (5) validation requirements for operational credibility.
• Evaluation Criteria: Studies are assessed qualitatively on dataset transparency, temporal/external validation, operational realism, constraint handling, and evidence of real-world deployment. The review explicitly avoids pooling effect sizes due to metric heterogeneity.

Key Contributions and Evidence Synthesis
The review organizes findings into distinct operational domains, highlighting the maturity and limitations of each:

1. Flight Delay Prediction: Methodological foundations are comparatively mature, utilizing tree ensembles, gradient boosting, and recurrent neural networks. However, the evidence base suffers from "translational weakness": models are often validated via random train/test splits on historical data, leading to temporal leakage and poor generalization to future schedules. Operational value requires temporal holdouts, calibration, and stability rather than just headline accuracy.
2. Delay Propagation: Research has moved beyond isolated flight prediction to network-based modeling (complex networks, graph neural networks, causal approaches). While these models effectively characterize how disturbances propagate through aircraft rotations and passenger connections, their transferability across different airlines and network structures remains a significant limitation.
3. Recovery (Aircraft, Crew, Passenger, Integrated):
   • Aircraft Recovery: Has a mature optimization foundation (MILP, column generation) but faces computational burdens and dependency on accurate state specification.
   • Crew Recovery: Highly constraint-intensive (legality, rest limits). The review emphasizes that any AI recommendation ignoring these hard constraints is operationally unusable, suggesting a need for hybrid architectures where AI prioritizes and optimization enforces feasibility.
   • Passenger & Integrated Recovery: Moving toward multi-resource optimization to prevent local optima from degrading other subsystems. However, full integration is hindered by data synchronization issues and organizational complexity.
4. AI and Reinforcement Learning (RL): RL is identified as an emerging direction for dynamic, sequential recovery. However, the review cautions that fast inference does not equate to operational superiority. Current evidence lacks sufficient stress testing, constraint verification, and comparison against strong optimization baselines to justify RL as a replacement for mathematical recovery.
5. Human-Centered Decision Support: The review posits that OCC deployment is a socio-technical problem. Systems must provide decision-centered explanations (e.g., specific drivers of risk), expose uncertainty, and preserve human authority. "Automation bias" and "alert fatigue" are key risks if interfaces are not designed for qualified operators.

Results and Critical Findings

• Prediction vs. Decision Value: A central finding is that predictive performance (e.g., AUC, RMSE) and operational decision value are distinct. A model can be highly accurate yet offer no operational benefit if the warning arrives too late or the suggested action is infeasible.
• The Prediction-to-Action Gap: The literature lacks robust evidence for end-to-end systems that successfully convert forecasts into feasible recovery actions in live environments.
• Validation Deficits: Most studies rely on internal validation (random splits) rather than temporal or external validation (future periods, different airlines). There is a scarcity of evidence regarding model drift, cybersecurity, and real-world "shadow-mode" evaluation.
• Maturity Hierarchy: Delay prediction and mathematical recovery are mature research areas; propagation modeling is increasingly sophisticated; integrated recovery is conceptually established but hard to operationalize; and AI/RL-based dynamic recovery with human-centered validation has a small evidence base.

Significance and Claims
The paper claims its primary significance lies in redefining the evaluation criteria for AI in airline operations. It argues that the field must shift from treating disruption management as a "prediction contest" to evaluating it as a decision-support problem.

• Operational Credibility: The review asserts that for AI to be operationally credible, it must demonstrate not just accuracy, but the ability to improve operational KPIs (cost, resilience, passenger outcomes) while maintaining safety, explainability, and human authority.
• Future Pathway: The paper proposes a staged validation pathway: moving from retrospective replay to shadow-mode deployment (where recommendations do not influence live operations) and finally to controlled decision-support trials. It concludes that the most credible near-term direction is hybrid and human-centered: using AI for situational awareness and optimization/rules for feasibility, with qualified OCC personnel retaining final decision authority.

The paper explicitly avoids claiming that current AI systems are ready for full autonomous deployment, emphasizing instead the need for rigorous external validation and the integration of uncertainty and human factors into the core design of disruption management systems.