Predicting Atmospheric Trace Substance Concentrations Using Supervised Machine Learning and HYSPLIT Backward Trajectories

Abstract Short-term meteorological influences are calculated from 24-h backward trajectories and are used as inputs into a supervised machine learning model with the goal of predicting airborne atmospheric trace substance (ATS) concentrations from the underlying meteorological signature. An introspective analysis shows that combining spatial reference features and meteorological data can capture a significant portion of the variability in ATS mixing ratios at—for some ATS—parts-per-trillion-level precision. This leads to the development of a machine learning ensemble system that combines predictions from a single global model with an ensemble of regional random forest and linear regression models. Overall, more than 40 000 samples from various airborne science campaigns spanning over 30 years are integrated into the model. The proposed model is then evaluated against a suite of chemical transport models, demonstrating that the ensemble system can produce explainable predictions while offering improved predictive ability compared to both coarse global models and high-resolution regional simulations evaluated at overlapping points. The model is then used to start a discussion on measurements of background atmospheric composition and the scope of their meteorological representation. Significance Statement Through knowledge of the prior meteorological conditions leading up to a measurement of atmospheric composition and its position on the globe, a model capable of predicting various atmospheric trace substances can be trained to accurately predict trace gas mixing ratios at high precision from a meteorological and spatial signature. This study’s approach can be used in interpolation tasks to fill spatial and temporal gaps in datasets with limited access to atmospheric chemistry data. However, significant challenges remain regarding model optimization for high signal-to-noise atmospheric compounds and for extrapolation tasks where predictions over entirely new spatial domains are needed.