Apportioning Firework Burning and Routine Source Contributions to PM_2.5 in a Chinese Megacity Using Two PMF-Machine Learning Frameworks

The firework burning (FB) during the Chinese Spring Festival can drastically elevate PM_2.5 concentrations and rapidly deteriorate local air quality, highlighting the need for timely quantification of its contributions to PM_2.5. However, as a short-term episodic source, FB poses substantial challenges for the accurate and rapid quantification of its contributions to PM_2.5. In this study, two source apportionment frameworks were developed that integrated the Positive Matrix Factorization (PMF) model with machine learning (ML), aiming to quantify the contributions of FB and routine sources to PM_2.5. The frameworks were applied to hourly measurements data of PM_2.5 chemical constituents collected in Shenzhen, China, from January to March 2023-2025. Twelve sources were identified, including 11 routine sources (such as secondary sulfate, secondary nitrate, vehicle emissions) and an episodic source of FB. Although the FB contributed only 4.7% to the total PM_2.5 mass over the entire period, its share surged to 33% during the firework display episodes and even exceeded 50% during peak burning hours. Two machine-learning-based approaches, the direct and indirect source apportionment methods, showed good fidelity to the PMF-resolved results and consistently identified FB as the dominant PM_2.5 source during firework display episodes, with estimated contributions comparable to those from the PMF model. The proposed PMF-ML framework captured episodic pollution signals while quantifying routine sources, with its ML component serving as a surrogate alternative to PMF.