Quantifying Aviation-Related Contributions to Ambient Ultrafine Particle Number Concentrations Using Interpretable Machine Learning

Ultrafine particles (UFP, D_p < 100 nm) are abundantly emitted by aircraft, but quantifying their contributions to ambient particle number concentrations (PNC) is challenging due to confounding from local traffic and complex interactions between aircraft plumes and meteorology. We applied a machine learning (ML) model to a multi-year PNC data set collected near Boston Logan International Airport, incorporating meteorology, road traffic, and runway-specific aircraft activity. We used SHapley Additive exPlanations (SHAP), a game-theoretic method that attributes feature contributions to model predictions, to interpret the black box ensemble ML model. SHAP enabled hourly source attribution, revealing feature interactions and nonlinear effects often missed by traditional tools (e.g., linear regression). The model performed well (R² = 0.66), exceeding typical hourly PNC models. SHAP results revealed that aircraft arrivals, particularly those on runways oriented perpendicular to the monitor-airport axis, were more influential than departures or on-ground airport activity. This suggests that aircraft not flying directly overhead can substantially impact ground-level air quality due to crosswinds. SHAP analysis further indicated that aircraft impacts depended on planetary boundary layer height, with intermediate heights associated with elevated PNC. This approach provides a novel and transferable framework for retrospective source-specific exposure assessment and improved characterization of aviation-related UFP in near-airport communities.