Predicting Aviation Contrail Occurrence Using Bayesian Population Statistics From Reanalysis Data

Abstract. Despite the ongoing climate crisis and recent pandemic-induced disruption, the aviation sector is expected to experience 5% annual growth over the next decade. While the industry moves towards decarbonisation through use of sustainable fuels and improved operating practices, the contribution by non-CO₂ effects become ever more apparent. Contrails and contrail-induced cirrus clouds contribute an estimated 57% to the sector's total effective radiative forcing (ERF). Contrail avoidance methods are gaining ground as tools to strategically reroute flights to reduce their ERF by predicting contrail forming regions in advance.

The task of prediction remains a challenge however, with typical methodologies employing either highly parametrised models that suffer from uncertainties, or machine learning methods that are heavily abstracted away from the background physics. We propose a novel, robust method for contrail prediction that leverages large-scale population behaviours. Using ERA-5 reanalysis and the OpenContrails dataset for over 50,000 confirmed contrails between 2019 and 2020 over North America, we train an informed contrail predictor using Bayesian methods which we verify on unseen data. Results and statistical evaluation of this model are presented, providing a scalable but interpretable contrail predictor with good skill (F₁ = 0.801) that could be run using output from numerical weather prediction models, or time-slice outputs from high-resolution climate models.