From receptor models to machine learning: Advancing source apportionment of carbonaceous aerosol

Abstract

Source apportionment of organic carbon (OC) and elemental carbon (EC) is essential for understanding combustion-related pollution, secondary aerosol formation, and their impacts on climate and human health. Conventional receptor models, such as Positive Matrix Factorization (PMF), remain widely used but are constrained by linear assumptions, sensitivity to input uncertainty, and limited ability to resolve overlapping and evolving sources. The integration of machine learning (ML) has emerged as a promising solution to these challenges, bridging the gap between linear statistical assumptions and the dynamic reality of ambient air quality monitoring. This review evaluates the role of ML in OC/EC source apportionment by organizing existing studies according to their functional roles within the workflow, including PMF-assisted hybrid approaches, data structuring methods, end-to-end frameworks, and physics-informed models. Particular attention is given to challenges related to data resolution, transferability, interpretability, and uncertainty quantification. The analysis shows that ML can help improve nonlinear source discrimination, high-resolution source prediction, and integration of heterogeneous datasets, yet most approaches remain dependent on receptor-model-derived labels and lack standardized validation protocols. Hybrid frameworks that combine receptor-model interpretability with ML predictive capability currently represent the most widely adopted approach. At the domain level, ML-assisted source apportionment can support improved characterization of emission sources and atmospheric processes, contributing to advances in aerosol chemistry and modeling. As a complementary tool that strengthens traditional receptor modeling, the integration of ML into source apportionment provides a more consistent evidence base for air quality policies and public health protection.