Data-driven modeling of environmental factors influencing Arctic methanesulfonic acid aerosol concentrations

Abstract. Natural aerosol components such as particulate methanesulfonic acid (MSA_p) play an important role in the Arctic climate. However, numerical models struggle to reproduce MSA_p concentrations and seasonality. Here we present an alternative data-driven methodology for modeling MSA_p at four High Arctic stations (Alert, Gruvebadet, Pituffik/Thule, and Utqiaġvik/Barrow). In our approach, we create input features that consider the ambient conditions during atmospheric transport (e.g., temperature, radiation, cloud cover, etc.) for use in two data-driven models: a random forest (RF) regressor and an additive model (AM). The most important features were selected through automatic selection procedures and their relationships with MSA_p model output was investigated. Although the overall performance of our data-driven models on test data is modest (max. R² = 0.29), the models can capture variability in the data well (max. Pearson correlation coefficient = 0.77), outperform the current numerical models and reanalysis products, and produce physically interpretable results.

The data-driven models selected features related to the sources, chemical processing, and removal of MSA_p with specific differences between stations. The seasonal cycles and selected features suggest gas-phase oxidation is relatively more important during peak concentration months at Alert, Gruvebadet, and Pituffik/Thule while aqueous-phase oxidation is relatively more important at Utqiaġvik/Barrow. Alert and Pituffik/Thule appear to be more influenced by processes aloft than in the boundary layer. Our models usually selected chemical processing related features as the main factors influencing MSA_p predictions, highlighting the importance of properly simulating oxidation related processes in numerical models.