Optimizing OVOCs Source Apportionment with Machine Learning-Enhanced Photochemical Age-based Parameterization Model

Oxygenated volatile organic compounds (OVOCs) are important precursors and intermediate products of atmospheric photochemical reactions, which can promote the formation of secondary pollutants such as ozone (O₃) and secondary organic aerosol (SOA). The photochemical age parameterization model is widely used to analyze primary and secondary sources of OVOCs. However, a key challenge lies in selecting appropriate tracers, chemicals used to estimate contributions from different emission sources. Accurate tracer selection is crucial for improving source apportionment accuracy, yet it is often constrained by local emission inventories and may not fully capture rapid atmospheric chemical transformations, introducing uncertainty in OVOC apportionment. This study presents a novel approach integrating eight different machine learning methods to identify optimal tracers for OVOCs during extreme summer temperatures (experimental group) and average spring temperatures (control group). Our results demonstrated notable differences in tracer effectiveness between these two groups. In the spring, toluene and carbon monoxide (CO) were identified as the most effective tracers for OVOCs with high and low reactivity, respectively. In the summer, acetylene or CO were better suited for moderate and low reactivity OVOCs. By incorporating machine learning for tracer selection, we significantly improved the accuracy of the photochemical age parameterization model. The machine learning outputs correlated well with the model’s performance, particularly in terms of fitting accuracy of OVOCs. However, extremely high temperatures during summer disrupted the usual patterns of OVOC production and removal, which led to inconsistencies in matching high reactivity OVOCs with their tracers. Future research involves collecting more data on OVOC behavior under high-temperature conditions and applying Fourier transformation techniques. This will help in identifying characteristic patterns and improving the dynamic accuracy of our model.