Estimating the atmospheric aerosol number size distribution using deep learning

Abstract

The submicron aerosol number size distribution significantly impacts human health, air quality, weather, and climate. However, its measurement requires sophisticated and expensive instrumentation that demands substantial maintenance efforts, leading to limited data availability. To tackle this challenge, we developed estimation models using advanced deep learning algorithms to estimate the aerosol number size distribution based on trace gas concentrations, meteorological parameters, and total aerosol number concentration. These models were trained and validated with 15 years of ambient data from three distinct environments, and data from a fourth station were exclusively used for testing. Our estimative models successfully replicated the trends in the test data, capturing the temporal variations of particles ranging from approximately 10–500 nm, and accurately deriving total number, surface area, and mass concentrations. The model's accuracy for particles below 75 nm is limited without the inclusion of total particle number concentration as training input, highlighting the importance of this parameter for capturing the dynamics of smaller particles. The reliance on total particle number concentration, a parameter not routinely measured at all in air quality monitoring sites, as a key input for accurate estimation of smaller particles presents a practical challenge for broader application of the models. Our models demonstrated a robust generalization capability, offering valuable data for health assessments, regional pollution studies, and climate modeling. The estimation models developed in this work are representative of ambient conditions in Finland, but the methodology in general can be applied in broader regions.