Classification and cross-site prediction of PM2.5 chemical composition patterns using a machine learning framework

This study presents an integrated machine learning framework for classifying PM 2.5 composition-based chemical regimes using the air quality management offices (AQMOs) database. By employing self-organizing maps (SOM) and K-means clustering, patterns in PM 2.5 composition were visualized by grouping observations with similar chemical characteristics. The identified regimes showed distinct composition-based features: urban and industrialized areas were associated with higher frequencies of industrial- and traffic-related chemical patterns, whereas coastal or island regions showed higher frequencies of sea-salt-related and background-like chemical regimes. Subsequently, the identified chemical regime labels were classified using a Random Forest model, achieving an overall accuracy of 0.921 and a macro F1-score of 0.843. The trained model was further applied to independent measurement sites (OS, HS, and GC) to evaluate cross-site applicability. The proposed framework provides a practical approach for characterizing PM 2.5 chemical regimes using multi-site observational data and may serve as a complementary preliminary diagnostic tool for small-scale studies or regions where detailed receptor-model results are not yet available.