Probabilistic Surrogate Modeling and Uncertainty Quantification for Aviation Fleet Emissions

The aviation sector’s climate impact has shown significant growth over recent years, and credible long-term forecasts are essential for planning sustainable transitions. However, most fleet-level studies rely on deterministic point estimates that obscure uncertainty from technology adoption, operational variability, and fuel pathway assumptions. This work introduces a probabilistic surrogate framework that integrates fleet-level simulation, deep ensemble modeling, and Bayesian uncertainty quantification to produce calibrated and interpretable forecasts of aviation emissions. A heteroscedastic deep ensemble surrogate is trained on outputs from the Interactive Dynamic Environment Analysis (IDEA) tool, which models fleet-wide fuel use under varying enabler configurations, to capture both epistemic (model-form) and aleatory (data-driven) uncertainty across multi-decade projections. Post-hoc variance scaling ensures well-calibrated predictive intervals whose empirical coverage matches nominal confidence levels. Results show that epistemic uncertainty decreases as data coverage expands, while aleatory uncertainty remains dominant, indicating that most variability in emissions originates from intrinsic operational and scenario-level noise rather than model insufficiency. The proposed approach delivers well-calibrated, statistically coherent forecasts and offers a scalable pathway for uncertainty-aware decision-making in fleet-level aviation climate impact assessments.