An Ensemble Machine Learning Approach for Predicting Sources of Organic Aerosols Measured by Aerosol Mass Spectrometry

Abstract

Long-term source apportionment of atmospheric organic aerosols (OA) is essential for supporting air pollution management strategies. While aerosol mass spectrometry (AMS) combined with traditional source apportionment tools can accurately identify various OA sources, they face efficiency challenges when processing large volumes of long-term data. This study proposes an ensemble machine learning approach to efficiently apportion OA sources from long-term AMS measurements. Using six-year observation of a simplified version of AMS (i.e., ACSM) in the Paris region along with OA factor data derived from positive matrix factorization analysis, we developed an ensemble machine learning source apportionment model. Compared to individual machine learning algorithms, the ensemble model substantially reduced the root-mean-square error (RMSE) and increased the correlation coefficient in predicting OA sources by approximately 30% and 5%, respectively. Sensitivity analysis with five years of baseline data revealed that model performance relatively stabilizes when the training data exceeds two years, with RMSE values for primary and secondary OA factors at 0.31–0.45 μg/m³ and 0.69–0.84 μg/m³, respectively. This ensemble model not only enhances the efficiency of long-term OA source apportionment but also holds potential for near-real-time online applications.