Data Quality and Predictive Reliability in AI-Enabled Aviation Energy Management: Physics-Informed Sensitivity Analysis with Open-Data Validation

The objective of this study is to quantify how independently degraded operational-data quality affects predictive accuracy, predictive uncertainty, and prediction-level eligibility for decision support in AI-enabled aviation energy management. The evidence base combines a 24,000-flight physics-informed synthetic experiment, an independent distribution-shift experiment, a 17,520 h synthetic airport-energy time series, and external validation using 686 observed fuel-burn intervals from 79 flights and 10 aircraft types in the EUROCONTROL PRC 2025/OpenSky dataset. In the primary flight experiment, 40% missingness increased Histogram Gradient Boosting (HGB) MAPE from 2.170% to 15.763%. In the airport experiment, 40% temporal/context misalignment increased peak-hour MAPE from 1.849% to 8.949%. Under distribution shift, nominal 90% conformal coverage fell to 73.39%; local recalibration restored coverage to 90.0% but reduced eligibility at the 15% relative-width gate from 69.06% to 18.40%. In the external validation, Random Forest and HGB achieved mean MAPE values of 25.02% and 25.43%, respectively, compared with 32.41% for the OpenAP physics-based baseline and 41.21% for the simple duration-rate baseline. The results show that data-quality dimensions are not interchangeable and that uncertainty-based eligibility thresholds require local calibration. The study establishes a predictive-reliability mechanism; it does not claim realized fuel or electricity savings. More broadly, the proposed methodology is relevant to green transformation and digital transformation in data-intensive engineering systems and offers transferable methodological insights for medical technologies, where reliable AI-supported decisions similarly depend on data quality, uncertainty calibration, and traceable decision processes.