New particle formation prediction by using a machine learning method

Abstract

New Particle Formation (NPF) is a major atmospheric process that impacts the climate, atmospheric chemistry, and human health. NPF events can occur anywhere around the world, and can make up to 50% of all aerosol number concentration. NPF events are considered a strong precursor for cloud condensation nuclei (CCN), which in turn affect air pollution and cloud formation. This study develops complex models to predict NPF events in the urban Gwangju site using meteorological and gas species information. Classified NPF days datasets from Gwangju, South Korea (2014-2025), along with temperature, pressure, relative humidity, solar radiation, cloud coverage, and various gas species concentrations, were implemented in machine learning (ML) models. Data augmentation was achieved with SMOTE, while feature selection was conducted by Kruskal-Wallis and Mutual Information (MI) tests. Results show that the Extreme Gradient Boosting (XGB) model illustrated the highest classification accuracy (81.82%) for the urban Gwangju site. The regression analysis performed on the models illustrated satisfactory results, with a moderate coefficient of determination (R2) score of 0.58 for the test dataset. Overall, basic features such as solar radiation, cloud coverage, temperature, heat index, wind speed, CO, and PM2.5 are observed to be effective, minimal predictors for the NPF events.