Accurate reconstruction of historical photolysis frequencies using a machine learning approach enables long-term photochemical analysis: Implications for future O3 mitigation.

The photolysis frequency (J value) is a critical parameter in atmospheric photochemical simulations. However, routine observations of J values remain scarce, introducing substantial uncertainties in ozone (O3) and photochemical analysis. In this study, significantly lower J values were observed at a typical urban site in the Fenwei Plain (FWP) region of China. J values simulated by the tropospheric ultraviolet and visible (TUV) radiative model were also significantly overestimated. We developed a machine learning model using routine air pollutant, meteorological, and reanalysis data aiming to accurately reconstruct long-term J values to constrain an observation-based model (OBM). The model showed great J value predictive performance (R2 = 0.92-0.93). The reconstructed historical J values exhibited interannual variations consistent with those of O3. Multi-year OBM simulations indicated that sustained emission reductions of nitrogen oxides (NOX) in recent years have driven a gradual shift in the O3 formation regime toward NOX-limited conditions, while pronounced seasonal variability has persisted. The primary production of the peroxyl radicals, rather than radical chain length (ChL), governed the monthly variation in O3 production. Alkenes, aromatics, and OVOCs were consistently the most sensitive VOC species for O3 formation, while the importance of CO increased markedly during autumn. Industrial and transportation sectors are major emission sources of O3 precursors. To shift to the NOX-sensitive regime in September and October, more than ∼54% and ∼75% reductions in NOX reactivity are needed. Overall, this study provides an efficient machine learning approach for reconstructing historical J values using routine observations, thereby facilitating long-term photochemical assessments and supporting air quality management in regions lacking routine J value observations.