Contrail Persistence Prediction Using a Probabilistic Deep Neural Network Based Hybrid Modeling Framework

Persistent contrails--ice clouds formed from aircraft emissions--can evolve into cirrus-shaped clouds that trap outgoing long-wave radiation, making substantial contribution to aviation-induced climate forcing. Although persistent contrails are transient, their radiative effects can rival or exceed that of aviation CO2; notably only 2–5% of flights are responsible for up to 80–95% of this warming, making targeted mitigation feasible. Existing trajectory-based prediction models for contrail-induced radiative effect trade fidelity for speed: fast trajectory-based tools like the Contrail Cirrus Prediction model (CoCiP) are simplified, while high-fidelity plume-scale process models such as the Aircraft Plume Chemistry, Emissions, and Microphysics Model (APCEMM) are too computationally intensive for real-time use. Bridging this gap is essential for Monitoring, Reporting and Verification of climate-aware flight route planning. We introduce HyCoP (Hybrid Contrail Prediction), a probabilistic deep-learning framework that fuses information from multiple data sources for contrail persistence prediction and builds an efficient surrogate while achieving accuracy of the highest fidelity data source. We demonstrate the performance of this framework by considering CoCiP and APCEMM data. A probabilistic deep neural network is first trained on large-scale CoCiP data, then fine-tuned with a smaller set of APCEMM simulations to inject high-fidelity chemistry-microphysics dynamics to contrail persistence. Inputs include aircraft and engine parameters at each flight waypoint and meteorological fields (dimension-reduced via principal component analysis) in local neighborhoods around each flight waypoint. HyCoP produces calibrated probabilistic predictions of contrail persistence that can be tuned to desired operational performance levels. We assess the trained HyCoP model and interpret its behavior using Sobol global sensitivity analysis, revealing the dominant influence of meteorological and aircraft-emission features. This work advances accurate, interpretable, and scalable tools for contrail mitigation and MRV.